<html xmlns="http://www.w3.org/1999/xhtml">
<head>
<meta charset="utf-8">
<meta name="generator" content="pdf2htmlEX">
<meta http-equiv="X-UA-Compatible" content="IE=edge,chrome=1">
<link rel="stylesheet" href="https://static.pudn.com/base/css/base.min.css">
<link rel="stylesheet" href="https://static.pudn.com/base/css/fancy.min.css">
<link rel="stylesheet" href="https://static.pudn.com/prod/directory_preview_static/625fb7df8cbeb85d57682fe6/raw.css">
<script src="https://static.pudn.com/base/js/compatibility.min.js"></script>
<script src="https://static.pudn.com/base/js/pdf2htmlEX.min.js"></script>
<script>
try{
pdf2htmlEX.defaultViewer = new pdf2htmlEX.Viewer({});
}catch(e){}
</script>
<title></title>
</head>
<body>
<div id="sidebar" style="display: none">
<div id="outline">
</div>
</div>
<div id="pf1" class="pf w0 h0" data-page-no="1"><div class="pc pc1 w0 h0"><img class="bi x0 y0 w1 h1" alt="" src="https://static.pudn.com/prod/directory_preview_static/625fb7df8cbeb85d57682fe6/bg1.jpg"><div class="t m0 x1 h2 y1 ff1 fs0 fc0 sc0 ls0 ws0"><span class="fc0 sc2">书</span><span class="_ _0"></span><span class="sc1"><span class="fc0 sc2">书</span><span class="_ _0"></span><span class="fc1 sc2"><span class="fc0 sc2">书</span></span></span></div><div class="t m0 x2 h3 y2 ff2 fs1 fc2 sc2 ls0 ws0">Ｉ<span class="_ _1"></span>Ｓ<span class="_ _2"></span>Ｓ<span class="_ _3"></span>Ｎ</div><div class="t m0 x3 h4 y3 ff3 fs1 fc2 sc2 ls0 ws0">　</div><div class="t m0 x4 h3 y2 ff2 fs1 fc2 sc2 ls0 ws0">１<span class="_ _4"></span>０<span class="_ _4"></span>０<span class="_ _4"></span>４</div><div class="t m1 x5 h3 y3 ff2 fs1 fc2 sc2 ls0 ws0">－</div><div class="t m0 x6 h3 y2 ff2 fs1 fc2 sc2 ls0 ws0">９<span class="_ _4"></span>０<span class="_ _4"></span>３<span class="_ _4"></span>７</div><div class="t m0 x7 h5 y4 ff4 fs1 fc2 sc2 ls0 ws0">，</div><div class="t m0 x8 h3 y2 ff2 fs1 fc2 sc2 ls0 ws0">Ｃ<span class="_ _5"></span>Ｏ<span class="_ _5"></span>Ｄ<span class="_ _5"></span>Ｅ<span class="_ _5"></span>Ｎ</div><div class="t m0 x9 h4 y3 ff3 fs1 fc2 sc2 ls0 ws0">　</div><div class="t m0 xa h3 y2 ff2 fs1 fc2 sc2 ls0 ws0">Ｓ<span class="_ _6"></span>Ｃ<span class="_ _5"></span>Ｙ<span class="_ _3"></span>Ｃ<span class="_ _3"></span>Ｅ<span class="_ _2"></span>４</div><div class="t m0 x2 h3 y5 ff2 fs1 fc2 sc2 ls0 ws0">Ｊ<span class="_ _1"></span>ｏ<span class="_ _4"></span>ｕ<span class="_ _1"></span>ｒ<span class="_ _1"></span>ｎ<span class="_ _4"></span>ａ<span class="_ _7"></span>ｌ</div><div class="t m0 xb h4 y6 ff3 fs1 fc2 sc2 ls0 ws0">　</div><div class="t m0 xc h3 y5 ff2 fs1 fc2 sc2 ls0 ws0">ｏ<span class="_ _8"></span>ｆ</div><div class="t m0 xd h4 y6 ff3 fs1 fc2 sc2 ls0 ws0">　</div><div class="t m0 x5 h3 y5 ff2 fs1 fc2 sc2 ls0 ws0">Ｄ<span class="_ _2"></span>ａ<span class="_ _8"></span>ｔ<span class="_ _8"></span>ａ</div><div class="t m0 xe h4 y6 ff3 fs1 fc2 sc2 ls0 ws0">　</div><div class="t m0 xf h3 y5 ff2 fs1 fc2 sc2 ls0 ws0">Ａ<span class="_ _6"></span>ｃ</div><div class="t m0 x10 h3 y7 ff2 fs1 fc2 sc2 ls0 ws0">ｑ</div><div class="t m0 x11 h3 y5 ff2 fs1 fc2 sc2 ls0 ws0">ｕ<span class="_ _7"></span>ｉ<span class="_ _7"></span>ｓ<span class="_ _9"></span>ｉ<span class="_ _9"></span>ｔ<span class="_ _a"></span>ｉ<span class="_ _7"></span>ｏ<span class="_ _4"></span>ｎ</div><div class="t m0 x12 h4 y6 ff3 fs1 fc2 sc2 ls0 ws0">　</div><div class="t m0 x13 h3 y5 ff2 fs1 fc2 sc2 ls0 ws0">ａ<span class="_ _4"></span>ｎ<span class="_ _4"></span>ｄ</div><div class="t m0 x14 h4 y6 ff3 fs1 fc2 sc2 ls0 ws0">　</div><div class="t m0 x15 h3 y5 ff2 fs1 fc2 sc2 ls0 ws0">Ｐ<span class="_ _4"></span>ｒ<span class="_ _1"></span>ｏ<span class="_ _1"></span>ｃ<span class="_ _8"></span>ｅ<span class="_ _1"></span>ｓ<span class="_ _8"></span>ｓ<span class="_ _9"></span>ｉ<span class="_ _7"></span>ｎ</div><div class="t m0 x16 h3 y7 ff2 fs1 fc2 sc2 ls0 ws0">ｇ</div><div class="t m0 x17 h4 y8 ff3 fs1 fc2 sc2 ls0 ws0">　</div><div class="t m0 x18 h3 y5 ff2 fs1 fc2 sc2 ls0 ws0">Ｖ<span class="_ _3"></span>ｏ<span class="_ _9"></span>ｌ<span class="_ _3"></span>．<span class="_ _b"></span>３<span class="_ _4"></span>２</div><div class="t m0 x19 h5 y9 ff4 fs1 fc2 sc2 ls0 ws0">，</div><div class="t m0 x1a h3 y5 ff2 fs1 fc2 sc2 ls0 ws0">Ｎ<span class="_ _3"></span>ｏ<span class="_ _5"></span>．<span class="_ _b"></span>２</div><div class="t m0 x1b h5 y9 ff4 fs1 fc2 sc2 ls0 ws0">，</div><div class="t m0 x1c h3 y5 ff2 fs1 fc2 sc2 ls0 ws0">Ｍ<span class="_ _3"></span>ａ<span class="_ _1"></span>ｒ<span class="_ _3"></span>．<span class="_ _b"></span>２<span class="_ _4"></span>０<span class="_ _4"></span>１<span class="_ _4"></span>７</div><div class="t m0 x1d h5 y9 ff4 fs1 fc2 sc2 ls0 ws0">，</div><div class="t m0 x1e h3 ya ff2 fs1 fc2 sc2 ls0 ws0">ｐ<span class="_ _4"></span>ｐ</div><div class="t m0 x1f h3 yb ff2 fs1 fc2 sc2 ls0 ws0">．<span class="_ _b"></span>２<span class="_ _4"></span>２<span class="_ _4"></span>１<span class="_ _0"></span>－<span class="_ _c"></span>２<span class="_ _4"></span>３<span class="_ _4"></span>１</div><div class="t m0 x20 h3 yc ff2 fs1 fc2 sc2 ls0 ws0">Ｄ<span class="_ _5"></span>Ｏ<span class="_ _6"></span>Ｉ</div><div class="t m0 x21 h5 yd ff4 fs1 fc2 sc2 ls0 ws0">：</div><div class="t m0 x22 h3 ye ff2 fs1 fc2 sc2 ls0 ws0">１<span class="_ _4"></span>０<span class="_ _c"></span>．<span class="_ _b"></span>１<span class="_ _4"></span>６<span class="_ _4"></span>３<span class="_ _4"></span>３<span class="_ _4"></span>７</div><div class="t m0 x23 h5 yd ff4 fs1 fc2 sc2 ls0 ws0">／</div><div class="t m0 xe h3 yf ff2 fs1 fc2 sc2 ls0 ws0">ｊ</div><div class="t m0 x24 h3 ye ff2 fs1 fc2 sc2 ls0 ws0">．<span class="_ _b"></span>１<span class="_ _4"></span>０<span class="_ _4"></span>０<span class="_ _4"></span>４</div><div class="t m1 x25 h3 y10 ff2 fs1 fc2 sc2 ls0 ws0">－</div><div class="t m0 x26 h3 ye ff2 fs1 fc2 sc2 ls0 ws0">９<span class="_ _4"></span>０<span class="_ _4"></span>３<span class="_ _4"></span>７<span class="_ _c"></span>．<span class="_ _b"></span>２<span class="_ _4"></span>０<span class="_ _4"></span>１<span class="_ _4"></span>７<span class="_ _0"></span>．<span class="_ _b"></span>０<span class="_ _4"></span>２<span class="_ _c"></span>．<span class="_ _b"></span>０<span class="_ _4"></span>０<span class="_ _4"></span>２</div><div class="t m0 x27 h6 y11 ff5 fs1 fc2 sc2 ls0 ws0"></div><div class="t m0 x28 h3 y12 ff2 fs1 fc2 sc2 ls0 ws0">２<span class="_ _4"></span>０<span class="_ _4"></span>１<span class="_ _4"></span>７<span class="_ _d"></span>ｂ</div><div class="t m0 x29 h3 y13 ff2 fs1 fc2 sc2 ls0 ws0">ｙ</div><div class="t m0 x5 h4 y14 ff3 fs1 fc2 sc2 ls0 ws0">　</div><div class="t m0 x6 h3 y12 ff2 fs1 fc2 sc2 ls0 ws0">Ｊ<span class="_ _1"></span>ｏ<span class="_ _4"></span>ｕ<span class="_ _1"></span>ｒ<span class="_ _1"></span>ｎ<span class="_ _4"></span>ａ<span class="_ _7"></span>ｌ</div><div class="t m0 x10 h4 y11 ff3 fs1 fc2 sc2 ls0 ws0">　</div><div class="t m0 x2a h3 y12 ff2 fs1 fc2 sc2 ls0 ws0">ｏ<span class="_ _8"></span>ｆ</div><div class="t m0 x2b h4 y11 ff3 fs1 fc2 sc2 ls0 ws0">　</div><div class="t m0 x2c h3 y12 ff2 fs1 fc2 sc2 ls0 ws0">Ｄ<span class="_ _2"></span>ａ<span class="_ _8"></span>ｔ<span class="_ _8"></span>ａ</div><div class="t m0 x12 h4 y11 ff3 fs1 fc2 sc2 ls0 ws0">　</div><div class="t m0 x13 h3 y12 ff2 fs1 fc2 sc2 ls0 ws0">Ａ<span class="_ _6"></span>ｃ</div><div class="t m0 x2d h3 y13 ff2 fs1 fc2 sc2 ls0 ws0">ｑ</div><div class="t m0 x14 h3 y12 ff2 fs1 fc2 sc2 ls0 ws0">ｕ<span class="_ _7"></span>ｉ<span class="_ _7"></span>ｓ<span class="_ _9"></span>ｉ<span class="_ _a"></span>ｔ<span class="_ _9"></span>ｉ<span class="_ _7"></span>ｏ<span class="_ _4"></span>ｎ</div><div class="t m0 x2e h4 y11 ff3 fs1 fc2 sc2 ls0 ws0">　</div><div class="t m0 x2f h3 y12 ff2 fs1 fc2 sc2 ls0 ws0">ａ<span class="_ _4"></span>ｎ<span class="_ _4"></span>ｄ</div><div class="t m0 x30 h4 y11 ff3 fs1 fc2 sc2 ls0 ws0">　</div><div class="t m0 x31 h3 y12 ff2 fs1 fc2 sc2 ls0 ws0">Ｐ<span class="_ _4"></span>ｒ<span class="_ _1"></span>ｏ<span class="_ _1"></span>ｃ<span class="_ _8"></span>ｅ<span class="_ _1"></span>ｓ<span class="_ _8"></span>ｓ<span class="_ _9"></span>ｉ<span class="_ _7"></span>ｎ</div><div class="t m0 x32 h3 y13 ff2 fs1 fc2 sc2 ls0 ws0">ｇ</div><div class="t m0 x33 h3 y2 ff2 fs1 fc2 sc2 ls0 ws0">ｈ<span class="_ _1"></span>ｔ<span class="_ _9"></span>ｔ</div><div class="t m0 x34 h3 y15 ff2 fs1 fc2 sc2 ls0 ws0">ｐ</div><div class="t m0 x35 h5 y4 ff4 fs1 fc2 sc2 ls0 ws0">：<span class="_ _b"></span>／<span class="_ _4"></span>／</div><div class="t m0 x36 h3 y2 ff2 fs1 fc2 sc2 ls0 ws0">ｓ</div><div class="t m0 x37 h3 y15 ff2 fs1 fc2 sc2 ls0 ws0">ｊ</div><div class="t m0 x38 h3 y2 ff2 fs1 fc2 sc2 ls0 ws0">ｃ</div><div class="t m0 x39 h3 y15 ff2 fs1 fc2 sc2 ls0 ws0">ｊ</div><div class="t m0 x3a h3 y2 ff2 fs1 fc2 sc2 ls0 ws0">．<span class="_ _a"></span>ｎ<span class="_ _2"></span>ｕ<span class="_ _4"></span>ａ<span class="_ _1"></span>ａ<span class="_ _5"></span>．<span class="_ _b"></span>ｅ<span class="_ _1"></span>ｄ<span class="_ _4"></span>ｕ<span class="_ _c"></span>．<span class="_ _b"></span>ｃ<span class="_ _4"></span>ｎ</div><div class="t m0 x3b h3 y5 ff2 fs1 fc2 sc2 ls0 ws0">Ｅ</div><div class="t m1 x3c h3 y6 ff2 fs1 fc2 sc2 ls0 ws0">－</div><div class="t m0 x3d h3 y5 ff2 fs1 fc2 sc2 ls0 ws0">ｍ<span class="_ _3"></span>ａ<span class="_ _9"></span>ｉ<span class="_ _a"></span>ｌ</div><div class="t m0 x3e h5 y9 ff4 fs1 fc2 sc2 ls0 ws0">：</div><div class="t m0 x3f h3 y5 ff2 fs1 fc2 sc2 ls0 ws0">ｓ</div><div class="t m0 x40 h3 y7 ff2 fs1 fc2 sc2 ls0 ws0">ｊ</div><div class="t m0 x41 h3 y5 ff2 fs1 fc2 sc2 ls0 ws0">ｃ</div><div class="t m0 x42 h3 y7 ff2 fs1 fc2 sc2 ls0 ws0">ｊ</div><div class="t m0 x43 h6 y6 ff5 fs1 fc2 sc2 ls0 ws0">＠</div><div class="t m0 x44 h3 y5 ff2 fs1 fc2 sc2 ls0 ws0">ｎ<span class="_ _4"></span>ｕ<span class="_ _4"></span>ａ<span class="_ _1"></span>ａ<span class="_ _5"></span>．<span class="_ _b"></span>ｅ<span class="_ _4"></span>ｄ<span class="_ _4"></span>ｕ<span class="_ _c"></span>．<span class="_ _b"></span>ｃ<span class="_ _1"></span>ｎ</div><div class="t m0 x45 h3 ye ff2 fs1 fc2 sc2 ls0 ws0">Ｔ<span class="_ _6"></span>ｅ<span class="_ _9"></span>ｌ</div><div class="t m0 x46 h5 yd ff4 fs1 fc2 sc2 ls0 ws0">／</div><div class="t m0 x3c h3 ye ff2 fs1 fc2 sc2 ls0 ws0">Ｆ<span class="_ _4"></span>ａ<span class="_ _4"></span>ｘ</div><div class="t m0 x35 h5 yd ff4 fs1 fc2 sc2 ls0 ws0">：</div><div class="t m0 x3e h3 ye ff2 fs1 fc2 sc2 ls0 ws0">＋<span class="_ _c"></span>８<span class="_ _4"></span>６</div><div class="t m1 x47 h3 y10 ff2 fs1 fc2 sc2 ls0 ws0">－</div><div class="t m0 x48 h3 ye ff2 fs1 fc2 sc2 ls0 ws0">０<span class="_ _4"></span>２<span class="_ _4"></span>５</div><div class="t m1 x49 h3 y10 ff2 fs1 fc2 sc2 ls0 ws0">－</div><div class="t m0 x4a h3 ye ff2 fs1 fc2 sc2 ls1 ws0">８４８９２７４２</div><div class="t m0 x4b h4 y11 ff3 fs1 fc2 sc2 ls0 ws0">　</div><div class="t m0 x27 h7 y16 ff6 fs2 fc2 sc2 ls2 ws0">基于深度学习的语音识别技术现状与展望</div><div class="t m2 x27 h8 y17 ff7 fs3 fc2 sc2 ls0 ws0">戴<span class="_ _e"></span>礼<span class="_ _e"></span>荣</div><div class="t m2 x4c h9 y18 ff3 fs3 fc2 sc2 ls0 ws0">　</div><div class="t m2 x4d h8 y17 ff7 fs3 fc2 sc2 ls0 ws0">张<span class="_ _e"></span>仕<span class="_ _e"></span>良</div><div class="t m2 x4e h9 y18 ff3 fs3 fc2 sc2 ls0 ws0">　</div><div class="t m2 x4f h8 y17 ff7 fs3 fc2 sc2 ls0 ws0">黄<span class="_ _e"></span>智<span class="_ _e"></span>颖</div><div class="t m0 x27 ha y19 ff4 fs4 fc2 sc2 ls0 ws0">（</div><div class="t m0 x50 hb y1a ff8 fs4 fc2 sc2 ls3 ws0">中国科学技术大学语音与语言信息处理国家工程实验室</div><div class="t m0 x51 ha y19 ff4 fs4 fc2 sc2 ls0 ws0">，</div><div class="t m0 x52 hb y1a ff8 fs4 fc2 sc2 ls0 ws0">合<span class="_ _f"></span>肥</div><div class="t m0 x53 ha y19 ff4 fs4 fc2 sc2 ls0 ws0">，</div><div class="t m0 x54 hc y1b ff2 fs4 fc2 sc2 ls0 ws0">２<span class="_ _1"></span>３<span class="_ _1"></span>０<span class="_ _1"></span>０<span class="_ _1"></span>２<span class="_ _1"></span>７</div><div class="t m0 x55 ha y19 ff4 fs4 fc2 sc2 ls0 ws0">）</div><div class="t m0 x27 hd y1c ff6 fs5 fc2 sc2 ls0 ws0">摘</div><div class="t m0 x56 he y1d ff3 fs5 fc2 sc2 ls0 ws0">　</div><div class="t m0 x57 hd y1c ff6 fs5 fc2 sc2 ls0 ws0">要</div><div class="t m0 x58 hf y1e ff9 fs5 fc2 sc2 ls0 ws0">：</div><div class="t m0 x59 hd y1c ffa fs5 fc2 sc2 ls4 ws0">首先对深度学习的发展历史以及概念进行简要的介绍</div><div class="t m0 x5a h10 y1e ffb fs5 fc2 sc2 ls0 ws0">。</div><div class="t m0 x5b hd y1c ffa fs5 fc2 sc2 ls5 ws0">然后<span class="_ _d"></span>回顾最近几年基于深度学习的语</div><div class="t m0 x27 hd y1f ffa fs5 fc2 sc2 ls4 ws0">音识别的研究进展</div><div class="t m0 x5c h10 y20 ffb fs5 fc2 sc2 ls0 ws0">。</div><div class="t m0 x12 hd y1f ffa fs5 fc2 sc2 ls4 ws0">这一部分内容主要分成以下</div><div class="t m0 x5d h11 y21 ff2 fs5 fc2 sc2 ls0 ws0">５</div><div class="t m0 x5e hd y1f ffa fs5 fc2 sc2 ls0 ws0">点<span class="_ _f"></span>进<span class="_ _f"></span>行<span class="_ _e"></span>介<span class="_ _f"></span>绍</div><div class="t m0 x5f h10 y20 ffb fs5 fc2 sc2 ls0 ws0">：</div><div class="t m0 x60 hd y1f ffa fs5 fc2 sc2 ls4 ws0">声学模型训练准则</div><div class="t m0 x61 h10 y20 ffb fs5 fc2 sc2 ls0 ws0">，</div><div class="t m0 x3d hd y1f ffa fs5 fc2 sc2 ls0 ws0">基<span class="_ _f"></span>于<span class="_ _f"></span>深<span class="_ _e"></span>度<span class="_ _e"></span>学<span class="_ _e"></span>习<span class="_ _e"></span>的<span class="_ _e"></span>声</div><div class="t m0 x27 hd y22 ffa fs5 fc2 sc2 ls0 ws0">学<span class="_ _f"></span>模<span class="_ _f"></span>型<span class="_ _f"></span>结<span class="_ _e"></span>构</div><div class="t m0 x62 h10 y23 ffb fs5 fc2 sc2 ls0 ws0">，</div><div class="t m0 x7 hd y24 ffa fs5 fc2 sc2 ls4 ws0">基于深度学习的声学模型训练效率优化</div><div class="t m0 x63 h10 y23 ffb fs5 fc2 sc2 ls0 ws0">，</div><div class="t m0 x64 hd y24 ffa fs5 fc2 sc2 ls4 ws0">基于深度学习的声学模型说话人自适应和基<span class="_ _10"></span>于深</div><div class="t m0 x27 hd y25 ffa fs5 fc2 sc2 ls4 ws0">度学习的端到端语音识别</div><div class="t m0 x65 h10 y26 ffb fs5 fc2 sc2 ls0 ws0">。</div><div class="t m0 x66 hd y25 ffa fs5 fc2 sc2 ls4 ws0">最后就基于深度学习的语音识别未来可能的研究方向进行展望</div><div class="t m0 x43 h10 y26 ffb fs5 fc2 sc2 ls0 ws0">。</div><div class="t m0 x27 hd y27 ff6 fs5 fc2 sc2 ls0 ws0">关<span class="_ _f"></span>键<span class="_ _f"></span>词</div><div class="t m0 x58 h10 y28 ffb fs5 fc2 sc2 ls0 ws0">：</div><div class="t m0 x59 hd y27 ffa fs5 fc2 sc2 ls0 ws0">深<span class="_ _f"></span>度<span class="_ _f"></span>学<span class="_ _e"></span>习</div><div class="t m0 x67 h10 y28 ffb fs5 fc2 sc2 ls0 ws0">；</div><div class="t m0 x68 hd y27 ffa fs5 fc2 sc2 ls0 ws0">深<span class="_ _f"></span>度<span class="_ _f"></span>神<span class="_ _e"></span>经<span class="_ _f"></span>网<span class="_ _f"></span>络</div><div class="t m0 x17 h10 y28 ffb fs5 fc2 sc2 ls0 ws0">；</div><div class="t m0 x69 hd y27 ffa fs5 fc2 sc2 ls0 ws0">语<span class="_ _f"></span>音<span class="_ _f"></span>识<span class="_ _e"></span>别</div><div class="t m0 x6a h10 y28 ffb fs5 fc2 sc2 ls0 ws0">；</div><div class="t m0 x1b hd y27 ffa fs5 fc2 sc2 ls0 ws0">说<span class="_ _f"></span>话<span class="_ _f"></span>人<span class="_ _e"></span>自<span class="_ _f"></span>适<span class="_ _f"></span>应</div><div class="t m0 x27 hd y29 ff6 fs5 fc2 sc2 ls0 ws0">中<span class="_ _f"></span>图<span class="_ _f"></span>分<span class="_ _e"></span>类<span class="_ _f"></span>号</div><div class="t m0 x62 h10 y2a ffb fs5 fc2 sc2 ls0 ws0">：</div><div class="t m0 x8 h11 y2b ff2 fs5 fc2 sc2 ls0 ws0">Ｔ<span class="_ _5"></span>Ｎ<span class="_ _4"></span>９<span class="_ _7"></span>１<span class="_ _8"></span>２<span class="_ _5"></span>．<span class="_ _11"></span>３</div><div class="t m0 x6b he y2c ff3 fs5 fc2 sc2 ls0 ws0">　<span class="_ _f"></span>　<span class="_ _f"></span>　</div><div class="t m0 x6c hd y29 ff6 fs5 fc2 sc2 ls0 ws0">文<span class="_ _f"></span>献<span class="_ _f"></span>标<span class="_ _e"></span>志<span class="_ _f"></span>码</div><div class="t m0 x32 h10 y2a ffb fs5 fc2 sc2 ls0 ws0">：</div><div class="t m0 x6d h11 y2b ff2 fs5 fc2 sc2 ls0 ws0">Ａ</div><div class="t m0 x6e h12 y2d ffc fs6 fc2 sc2 ls0 ws0">Ｄ<span class="_ _9"></span>ｅ<span class="_ _12"></span>ｅ</div><div class="t m0 x3 h12 y2e ffc fs6 fc2 sc2 ls0 ws0">ｐ</div><div class="t m0 xb h13 y2f ff3 fs6 fc2 sc2 ls0 ws0">　</div><div class="t m0 x6f h12 y2d ffc fs6 fc2 sc2 ls0 ws0">Ｌ<span class="_ _a"></span>ｅ<span class="_ _13"></span>ａ<span class="_ _13"></span>ｒ<span class="_ _13"></span>ｎ<span class="_ _12"></span>ｉ<span class="_ _13"></span>ｎ</div><div class="t m0 x2a h12 y2e ffc fs6 fc2 sc2 ls0 ws0">ｇ</div><div class="t m0 x70 h13 y2f ff3 fs6 fc2 sc2 ls0 ws0">　</div><div class="t m0 x4e h12 y2d ffc fs6 fc2 sc2 ls0 ws0">ｆ<span class="_ _12"></span>ｏ<span class="_ _13"></span>ｒ</div><div class="t m0 x71 h13 y30 ff3 fs6 fc2 sc2 ls0 ws0">　</div><div class="t m0 x13 h12 y2d ffc fs6 fc2 sc2 ls0 ws0">Ｓ</div><div class="t m0 x72 h12 y2e ffc fs6 fc2 sc2 ls0 ws0">ｐ</div><div class="t m0 x73 h12 y2d ffc fs6 fc2 sc2 ls0 ws0">ｅ<span class="_ _12"></span>ｅ<span class="_ _12"></span>ｃ<span class="_ _b"></span>ｈ</div><div class="t m0 x6c h13 y30 ff3 fs6 fc2 sc2 ls0 ws0">　</div><div class="t m0 x74 h12 y2d ffc fs6 fc2 sc2 ls0 ws0">Ｒ<span class="_ _9"></span>ｅ<span class="_ _12"></span>ｃ<span class="_ _13"></span>ｏ</div><div class="t m0 x75 h12 y2e ffc fs6 fc2 sc2 ls0 ws0">ｇ</div><div class="t m0 x76 h12 y2d ffc fs6 fc2 sc2 ls0 ws0">ｎ<span class="_ _12"></span>ｉ<span class="_ _14"></span>ｔ<span class="_ _14"></span>ｉ<span class="_ _11"></span>ｏ<span class="_ _b"></span>ｎ</div><div class="t m0 x77 h14 y31 ff4 fs6 fc2 sc2 ls0 ws0">：</div><div class="t m0 x78 h12 y2d ffc fs6 fc2 sc2 ls0 ws0">Ｒ<span class="_ _9"></span>ｅ<span class="_ _13"></span>ｖ<span class="_ _15"></span>ｉ<span class="_ _15"></span>ｅ<span class="_ _9"></span>ｗ</div><div class="t m0 x1e h13 y30 ff3 fs6 fc2 sc2 ls0 ws0">　</div><div class="t m0 x79 h12 y2d ffc fs6 fc2 sc2 ls0 ws0">ｏ<span class="_ _12"></span>ｆ</div><div class="t m0 x2 h12 y32 ffc fs6 fc2 sc2 ls0 ws0">Ｓ<span class="_ _12"></span>ｔ<span class="_ _12"></span>ａ<span class="_ _12"></span>ｔ<span class="_ _15"></span>ｅ</div><div class="t m3 xb h15 y33 ff2 fs6 fc2 sc2 ls0 ws0">－</div><div class="t m0 xc h12 y34 ffc fs6 fc2 sc2 ls0 ws0">ｏ<span class="_ _12"></span>ｆ</div><div class="t m3 x6 h15 y33 ff2 fs6 fc2 sc2 ls0 ws0">－</div><div class="t m0 x7a h12 y34 ffc fs6 fc2 sc2 ls0 ws0">ｔ<span class="_ _13"></span>ｈ<span class="_ _13"></span>ｅ</div><div class="t m3 x8 h15 y33 ff2 fs6 fc2 sc2 ls0 ws0">－</div><div class="t m0 x7b h12 y34 ffc fs6 fc2 sc2 ls0 ws0">Ａ<span class="_ _9"></span>ｒ<span class="_ _15"></span>ｔ<span class="_ _15"></span>ｓ</div><div class="t m0 x7c h13 y33 ff3 fs6 fc2 sc2 ls0 ws0">　</div><div class="t m0 x9 h12 y34 ffc fs6 fc2 sc2 ls0 ws0">Ｔ<span class="_ _b"></span>ｅ<span class="_ _13"></span>ｃ<span class="_ _13"></span>ｈ<span class="_ _9"></span>ｎ<span class="_ _b"></span>ｏ<span class="_ _11"></span>ｌ<span class="_ _11"></span>ｏ</div><div class="t m0 x7d h12 y35 ffc fs6 fc2 sc2 ls0 ws0">ｇ</div><div class="t m0 x7e h12 y34 ffc fs6 fc2 sc2 ls0 ws0">ｉ<span class="_ _15"></span>ｅ<span class="_ _11"></span>ｓ</div><div class="t m0 x31 h13 y33 ff3 fs6 fc2 sc2 ls0 ws0">　</div><div class="t m0 x7f h12 y34 ffc fs6 fc2 sc2 ls0 ws0">ａ<span class="_ _b"></span>ｎ<span class="_ _a"></span>ｄ</div><div class="t m0 x80 h13 y33 ff3 fs6 fc2 sc2 ls0 ws0">　</div><div class="t m0 x81 h12 y34 ffc fs6 fc2 sc2 ls0 ws0">Ｐ<span class="_ _b"></span>ｒ<span class="_ _13"></span>ｏ<span class="_ _13"></span>ｓ</div><div class="t m0 x82 h12 y35 ffc fs6 fc2 sc2 ls0 ws0">ｐ</div><div class="t m0 x83 h12 y34 ffc fs6 fc2 sc2 ls0 ws0">ｅ<span class="_ _12"></span>ｃ<span class="_ _11"></span>ｔ<span class="_ _15"></span>ｓ</div><div class="t m0 x6e h11 y36 ff2 fs5 fc2 sc2 ls0 ws0">Ｄ<span class="_ _1"></span>ａ<span class="_ _b"></span>ｉ</div><div class="t m0 x84 he y37 ff3 fs5 fc2 sc2 ls0 ws0">　</div><div class="t m0 x3 h11 y36 ff2 fs5 fc2 sc2 ls0 ws0">Ｌ<span class="_ _9"></span>ｉ<span class="_ _13"></span>ｒ<span class="_ _a"></span>ｏ<span class="_ _8"></span>ｎ</div><div class="t m0 x85 h11 y38 ff2 fs5 fc2 sc2 ls0 ws0">ｇ</div><div class="t m0 x86 h16 y39 ff4 fs5 fc2 sc2 ls0 ws0">，</div><div class="t m0 xe h11 y36 ff2 fs5 fc2 sc2 ls0 ws0">Ｚ<span class="_ _1"></span>ｈ<span class="_ _7"></span>ａ<span class="_ _7"></span>ｎ</div><div class="t m0 x87 h11 y38 ff2 fs5 fc2 sc2 ls0 ws0">ｇ</div><div class="t m0 x26 he y3a ff3 fs5 fc2 sc2 ls0 ws0">　</div><div class="t m0 x5c h11 y36 ff2 fs5 fc2 sc2 ls0 ws0">Ｓ<span class="_ _8"></span>ｈ<span class="_ _a"></span>ｉ<span class="_ _11"></span>ｌ<span class="_ _12"></span>ｉ<span class="_ _13"></span>ａ<span class="_ _8"></span>ｎ</div><div class="t m0 x15 h11 y38 ff2 fs5 fc2 sc2 ls0 ws0">ｇ</div><div class="t m0 x88 h16 y39 ff4 fs5 fc2 sc2 ls0 ws0">，</div><div class="t m0 x89 h11 y36 ff2 fs5 fc2 sc2 ls0 ws0">Ｈ<span class="_ _3"></span>ｕ<span class="_ _9"></span>ａ<span class="_ _7"></span>ｎ</div><div class="t m0 x17 h11 y38 ff2 fs5 fc2 sc2 ls0 ws0">ｇ</div><div class="t m0 x8a he y3a ff3 fs5 fc2 sc2 ls0 ws0">　</div><div class="t m0 x8b h11 y36 ff2 fs5 fc2 sc2 ls0 ws0">Ｚ<span class="_ _1"></span>ｈ<span class="_ _b"></span>ｉ</div><div class="t m0 x8c h11 y38 ff2 fs5 fc2 sc2 ls0 ws0">ｙ</div><div class="t m0 x80 h11 y36 ff2 fs5 fc2 sc2 ls0 ws0">ｉ<span class="_ _b"></span>ｎ</div><div class="t m0 x32 h11 y38 ff2 fs5 fc2 sc2 ls0 ws0">ｇ</div><div class="t m0 x27 h5 y3b ff4 fs1 fc2 sc2 ls0 ws0">（</div><div class="t m0 x8d h3 y3c ff2 fs1 fc2 sc2 ls0 ws0">Ｎ<span class="_ _6"></span>ａ<span class="_ _8"></span>ｔ<span class="_ _9"></span>ｉ<span class="_ _7"></span>ｏ<span class="_ _4"></span>ｎ<span class="_ _4"></span>ａ<span class="_ _9"></span>ｌ</div><div class="t m0 x29 h4 y3d ff3 fs1 fc2 sc2 ls0 ws0">　</div><div class="t m0 x8e h3 y3c ff2 fs1 fc2 sc2 ls0 ws0">Ｅ<span class="_ _6"></span>ｎ</div><div class="t m0 x4d h3 y3e ff2 fs1 fc2 sc2 ls0 ws0">ｇ</div><div class="t m0 x8f h3 y3c ff2 fs1 fc2 sc2 ls0 ws0">ｉ<span class="_ _7"></span>ｎ<span class="_ _4"></span>ｅ<span class="_ _1"></span>ｅ<span class="_ _8"></span>ｒ<span class="_ _9"></span>ｉ<span class="_ _7"></span>ｎ</div><div class="t m0 x70 h3 y3e ff2 fs1 fc2 sc2 ls0 ws0">ｇ</div><div class="t m0 x4e h4 y3f ff3 fs1 fc2 sc2 ls0 ws0">　</div><div class="t m0 x90 h3 y3c ff2 fs1 fc2 sc2 ls0 ws0">Ｌ<span class="_ _2"></span>ａ<span class="_ _4"></span>ｂ<span class="_ _4"></span>ｏ<span class="_ _1"></span>ｒ<span class="_ _8"></span>ａ<span class="_ _8"></span>ｔ<span class="_ _8"></span>ｏ<span class="_ _1"></span>ｒ</div><div class="t m0 x91 h3 y3e ff2 fs1 fc2 sc2 ls0 ws0">ｙ</div><div class="t m0 x92 h4 y3f ff3 fs1 fc2 sc2 ls0 ws0">　</div><div class="t m0 x93 h3 y3c ff2 fs1 fc2 sc2 ls0 ws0">ｏ<span class="_ _8"></span>ｆ</div><div class="t m0 x89 h4 y3d ff3 fs1 fc2 sc2 ls0 ws0">　</div><div class="t m0 x94 h3 y3c ff2 fs1 fc2 sc2 ls0 ws0">Ｓ</div><div class="t m0 x7d h3 y3e ff2 fs1 fc2 sc2 ls0 ws0">ｐ</div><div class="t m0 x7e h3 y3c ff2 fs1 fc2 sc2 ls0 ws0">ｅ<span class="_ _1"></span>ｅ<span class="_ _8"></span>ｃ<span class="_ _4"></span>ｈ</div><div class="t m0 x95 h4 y3d ff3 fs1 fc2 sc2 ls0 ws0">　</div><div class="t m0 x8a h3 y3c ff2 fs1 fc2 sc2 ls0 ws0">ａ<span class="_ _4"></span>ｎ<span class="_ _4"></span>ｄ</div><div class="t m0 x96 h4 y3d ff3 fs1 fc2 sc2 ls0 ws0">　</div><div class="t m0 x97 h3 y3c ff2 fs1 fc2 sc2 ls0 ws0">Ｌ<span class="_ _2"></span>ａ<span class="_ _4"></span>ｎ</div><div class="t m0 x32 h3 y3e ff2 fs1 fc2 sc2 ls0 ws0">ｇ</div><div class="t m0 x6d h3 y3c ff2 fs1 fc2 sc2 ls0 ws0">ｕ<span class="_ _4"></span>ａ</div><div class="t m0 x98 h3 y3e ff2 fs1 fc2 sc2 ls0 ws0">ｇ</div><div class="t m0 x1b h3 y3c ff2 fs1 fc2 sc2 ls0 ws0">ｅ</div><div class="t m0 x99 h4 y3d ff3 fs1 fc2 sc2 ls0 ws0">　</div><div class="t m0 x9a h3 y3c ff2 fs1 fc2 sc2 ls0 ws0">Ｉ<span class="_ _1"></span>ｎ<span class="_ _8"></span>ｆ<span class="_ _8"></span>ｏ<span class="_ _1"></span>ｒ<span class="_ _6"></span>ｍ<span class="_ _3"></span>ａ<span class="_ _8"></span>ｔ<span class="_ _9"></span>ｉ<span class="_ _7"></span>ｏ<span class="_ _4"></span>ｎ</div><div class="t m0 x9b h4 y3d ff3 fs1 fc2 sc2 ls0 ws0">　</div><div class="t m0 x53 h3 y3c ff2 fs1 fc2 sc2 ls0 ws0">Ｐ<span class="_ _4"></span>ｒ<span class="_ _1"></span>ｏ<span class="_ _1"></span>ｃ<span class="_ _8"></span>ｅ<span class="_ _1"></span>ｓ<span class="_ _8"></span>ｓ<span class="_ _9"></span>ｉ<span class="_ _7"></span>ｎ</div><div class="t m0 x55 h3 y3e ff2 fs1 fc2 sc2 ls0 ws0">ｇ</div><div class="t m0 x5b h5 y3b ff4 fs1 fc2 sc2 ls0 ws0">，</div><div class="t m0 x9c h3 y3c ff2 fs1 fc2 sc2 ls0 ws0">Ｕ<span class="_ _3"></span>ｎ<span class="_ _8"></span>ｉ<span class="_ _7"></span>ｖ<span class="_ _1"></span>ｅ<span class="_ _1"></span>ｒ<span class="_ _7"></span>ｓ<span class="_ _9"></span>ｉ<span class="_ _9"></span>ｔ</div><div class="t m0 x9d h3 y3e ff2 fs1 fc2 sc2 ls0 ws0">ｙ</div><div class="t m0 x9e h4 y3f ff3 fs1 fc2 sc2 ls0 ws0">　</div><div class="t m0 x9f h3 y3c ff2 fs1 fc2 sc2 ls0 ws0">ｏ<span class="_ _8"></span>ｆ</div><div class="t m0 x6e h3 y40 ff2 fs1 fc2 sc2 ls0 ws0">Ｓ<span class="_ _1"></span>ｃ<span class="_ _9"></span>ｉ<span class="_ _7"></span>ｅ<span class="_ _4"></span>ｎ<span class="_ _1"></span>ｃ<span class="_ _1"></span>ｅ</div><div class="t m0 xb h4 y41 ff3 fs1 fc2 sc2 ls0 ws0">　</div><div class="t m0 xa0 h3 y40 ff2 fs1 fc2 sc2 ls0 ws0">ａ<span class="_ _4"></span>ｎ<span class="_ _4"></span>ｄ</div><div class="t m0 x85 h4 y41 ff3 fs1 fc2 sc2 ls0 ws0">　</div><div class="t m0 xa1 h3 y40 ff2 fs1 fc2 sc2 ls0 ws0">Ｔ<span class="_ _6"></span>ｅ<span class="_ _8"></span>ｃ<span class="_ _4"></span>ｈ<span class="_ _4"></span>ｎ<span class="_ _4"></span>ｏ<span class="_ _8"></span>ｌ<span class="_ _7"></span>ｏ</div><div class="t m0 xa2 h3 y42 ff2 fs1 fc2 sc2 ls0 ws0">ｇ<span class="_ _4"></span>ｙ</div><div class="t m0 x67 h4 y43 ff3 fs1 fc2 sc2 ls0 ws0">　</div><div class="t m0 x7c h3 y40 ff2 fs1 fc2 sc2 ls0 ws0">ｏ<span class="_ _8"></span>ｆ</div><div class="t m0 xa3 h4 y41 ff3 fs1 fc2 sc2 ls0 ws0">　</div><div class="t m0 x12 h3 y40 ff2 fs1 fc2 sc2 ls0 ws0">Ｃ<span class="_ _6"></span>ｈ<span class="_ _7"></span>ｉ<span class="_ _8"></span>ｎ<span class="_ _4"></span>ａ</div><div class="t m0 xa4 h5 y44 ff4 fs1 fc2 sc2 ls0 ws0">，</div><div class="t m0 xa5 h3 y40 ff2 fs1 fc2 sc2 ls0 ws0">Ｈ<span class="_ _3"></span>ｅ<span class="_ _7"></span>ｆ<span class="_ _8"></span>ｅ<span class="_ _9"></span>ｉ</div><div class="t m0 xa6 h5 y44 ff4 fs1 fc2 sc2 ls0 ws0">，</div><div class="t m0 xa7 h3 y40 ff2 fs1 fc2 sc2 ls0 ws0">２<span class="_ _4"></span>３<span class="_ _4"></span>０<span class="_ _4"></span>０<span class="_ _4"></span>２<span class="_ _4"></span>７</div><div class="t m0 xa8 h5 y44 ff4 fs1 fc2 sc2 ls0 ws0">，</div><div class="t m0 xa9 h3 y40 ff2 fs1 fc2 sc2 ls0 ws0">Ｃ<span class="_ _6"></span>ｈ<span class="_ _7"></span>ｉ<span class="_ _8"></span>ｎ<span class="_ _4"></span>ａ</div><div class="t m0 x77 h5 y44 ff4 fs1 fc2 sc2 ls0 ws0">）</div><div class="t m0 x6e h17 y45 ffc fs5 fc2 sc2 ls0 ws0">Ａ<span class="_ _4"></span>ｂ<span class="_ _a"></span>ｓ<span class="_ _13"></span>ｔ<span class="_ _b"></span>ｒ<span class="_ _a"></span>ａ<span class="_ _9"></span>ｃ<span class="_ _b"></span>ｔ</div><div class="t m0 x8e h16 y46 ff4 fs5 fc2 sc2 ls0 ws0">：</div><div class="t m0 xaa h11 y45 ff2 fs5 fc2 sc2 ls0 ws0">Ｉ<span class="_ _9"></span>ｎ</div><div class="t m0 x23 he y47 ff3 fs5 fc2 sc2 ls0 ws0">　</div><div class="t m0 xe h11 y45 ff2 fs5 fc2 sc2 ls0 ws0">ｔ<span class="_ _9"></span>ｈ<span class="_ _b"></span>ｉ<span class="_ _13"></span>ｓ</div><div class="t m0 xab he y47 ff3 fs5 fc2 sc2 ls0 ws0">　</div><div class="t m0 x87 h11 y48 ff2 fs5 fc2 sc2 ls0 ws0">ｐ</div><div class="t m0 x90 h11 y45 ff2 fs5 fc2 sc2 ls0 ws0">ａ</div><div class="t m0 x5c h11 y48 ff2 fs5 fc2 sc2 ls0 ws0">ｐ</div><div class="t m0 xac h11 y45 ff2 fs5 fc2 sc2 ls0 ws0">ｅ<span class="_ _a"></span>ｒ</div><div class="t m0 xad h16 y46 ff4 fs5 fc2 sc2 ls0 ws0">，</div><div class="t m0 xae h11 y45 ff2 fs5 fc2 sc2 ls0 ws0">ｄ<span class="_ _9"></span>ｅ<span class="_ _9"></span>ｅ</div><div class="t m0 x93 h11 y48 ff2 fs5 fc2 sc2 ls0 ws0">ｐ</div><div class="t m0 x66 he y49 ff3 fs5 fc2 sc2 ls0 ws0">　</div><div class="t m0 xaf h11 y45 ff2 fs5 fc2 sc2 ls0 ws0">ｌ<span class="_ _b"></span>ｅ<span class="_ _a"></span>ａ<span class="_ _a"></span>ｒ<span class="_ _7"></span>ｎ<span class="_ _b"></span>ｉ<span class="_ _b"></span>ｎ</div><div class="t m0 x8b h11 y48 ff2 fs5 fc2 sc2 ls0 ws0">ｇ</div><div class="t m0 xb0 he y49 ff3 fs5 fc2 sc2 ls0 ws0">　</div><div class="t m0 xb1 h11 y45 ff2 fs5 fc2 sc2 ls0 ws0">ｉ<span class="_ _13"></span>ｓ</div><div class="t m0 xb2 he y47 ff3 fs5 fc2 sc2 ls0 ws0">　</div><div class="t m0 xa9 h11 y45 ff2 fs5 fc2 sc2 ls0 ws0">ｂ<span class="_ _9"></span>ｒ<span class="_ _13"></span>ｉ<span class="_ _13"></span>ｅ<span class="_ _b"></span>ｆ<span class="_ _13"></span>ｌ</div><div class="t m0 xb3 h11 y48 ff2 fs5 fc2 sc2 ls0 ws0">ｙ</div><div class="t m0 xb4 he y49 ff3 fs5 fc2 sc2 ls0 ws0">　</div><div class="t m0 x83 h11 y45 ff2 fs5 fc2 sc2 ls0 ws0">ｉ<span class="_ _b"></span>ｎ<span class="_ _9"></span>ｔ<span class="_ _b"></span>ｒ<span class="_ _a"></span>ｏ<span class="_ _7"></span>ｄ<span class="_ _8"></span>ｕ<span class="_ _7"></span>ｃ<span class="_ _9"></span>ｅ<span class="_ _9"></span>ｄ<span class="_ _3"></span>．<span class="_ _c"></span>Ｔ<span class="_ _2"></span>ｈ<span class="_ _7"></span>ｅ<span class="_ _7"></span>ｎ</div><div class="t m0 x5b h16 y46 ff4 fs5 fc2 sc2 ls0 ws0">，</div><div class="t m0 x9c h11 y45 ff2 fs5 fc2 sc2 ls0 ws0">ａ</div><div class="t m0 xb5 he y47 ff3 fs5 fc2 sc2 ls0 ws0">　</div><div class="t m0 xb6 h11 y45 ff2 fs5 fc2 sc2 ls0 ws0">ｒ<span class="_ _a"></span>ｅ<span class="_ _9"></span>ｖ<span class="_ _b"></span>ｉ<span class="_ _13"></span>ｅ<span class="_ _4"></span>ｗ</div><div class="t m0 xb7 he y47 ff3 fs5 fc2 sc2 ls0 ws0">　</div><div class="t m0 xb8 h11 y45 ff2 fs5 fc2 sc2 ls0 ws0">ｏ<span class="_ _b"></span>ｆ</div><div class="t m0 xb9 he y47 ff3 fs5 fc2 sc2 ls0 ws0">　</div><div class="t m0 xba h11 y45 ff2 fs5 fc2 sc2 ls0 ws0">ｔ<span class="_ _9"></span>ｈ<span class="_ _7"></span>ｅ</div><div class="t m0 xbb he y47 ff3 fs5 fc2 sc2 ls0 ws0">　</div><div class="t m0 xbc h11 y45 ff2 fs5 fc2 sc2 ls0 ws0">ｒ<span class="_ _a"></span>ｅ<span class="_ _9"></span>ｓ<span class="_ _a"></span>ｅ<span class="_ _9"></span>ａ<span class="_ _a"></span>ｒ<span class="_ _a"></span>ｃ<span class="_ _7"></span>ｈ</div><div class="t m0 xbd he y47 ff3 fs5 fc2 sc2 ls0 ws0">　</div><div class="t m0 xbe h11 y48 ff2 fs5 fc2 sc2 ls0 ws0">ｐ</div><div class="t m0 x44 h11 y45 ff2 fs5 fc2 sc2 ls0 ws0">ｒ<span class="_ _9"></span>ｏ</div><div class="t m0 x49 h11 y48 ff2 fs5 fc2 sc2 ls0 ws0">ｇ</div><div class="t m0 xbf h11 y45 ff2 fs5 fc2 sc2 ls0 ws0">ｒ<span class="_ _a"></span>ｅ<span class="_ _9"></span>ｓ<span class="_ _a"></span>ｓ</div><div class="t m0 xc0 he y47 ff3 fs5 fc2 sc2 ls0 ws0">　</div><div class="t m0 xc1 h11 y45 ff2 fs5 fc2 sc2 ls0 ws0">ｏ<span class="_ _b"></span>ｆ</div><div class="t m0 x2 h11 y4a ff2 fs5 fc2 sc2 ls0 ws0">ｄ<span class="_ _9"></span>ｅ<span class="_ _9"></span>ｅ</div><div class="t m0 xc2 h11 y4b ff2 fs5 fc2 sc2 ls0 ws0">ｐ</div><div class="t m0 x3 he y4c ff3 fs5 fc2 sc2 ls0 ws0">　</div><div class="t m0 xc3 h11 y4a ff2 fs5 fc2 sc2 ls0 ws0">ｌ<span class="_ _b"></span>ｅ<span class="_ _a"></span>ａ<span class="_ _a"></span>ｒ<span class="_ _7"></span>ｎ<span class="_ _b"></span>ｉ<span class="_ _b"></span>ｎ</div><div class="t m0 x23 h11 y4b ff2 fs5 fc2 sc2 ls0 ws0">ｇ</div><div class="t m0 xf he y4c ff3 fs5 fc2 sc2 ls0 ws0">　</div><div class="t m0 x24 h11 y4a ff2 fs5 fc2 sc2 ls0 ws0">ｂ<span class="_ _7"></span>ａ<span class="_ _9"></span>ｓ<span class="_ _a"></span>ｅ<span class="_ _7"></span>ｄ</div><div class="t m0 xc4 he y4d ff3 fs5 fc2 sc2 ls0 ws0">　</div><div class="t m0 xc5 h11 y4a ff2 fs5 fc2 sc2 ls0 ws0">ｓ</div><div class="t m0 xa h11 y4b ff2 fs5 fc2 sc2 ls0 ws0">ｐ</div><div class="t m0 xc6 h11 y4a ff2 fs5 fc2 sc2 ls0 ws0">ｅ<span class="_ _9"></span>ｅ<span class="_ _a"></span>ｃ<span class="_ _7"></span>ｈ</div><div class="t m0 xc7 he y4d ff3 fs5 fc2 sc2 ls0 ws0">　</div><div class="t m0 x92 h11 y4a ff2 fs5 fc2 sc2 ls0 ws0">ｒ<span class="_ _a"></span>ｅ<span class="_ _9"></span>ｃ<span class="_ _9"></span>ｏ</div><div class="t m0 xc8 h11 y4b ff2 fs5 fc2 sc2 ls0 ws0">ｇ</div><div class="t m0 xc9 h11 y4a ff2 fs5 fc2 sc2 ls0 ws0">ｎ<span class="_ _b"></span>ｉ<span class="_ _12"></span>ｔ<span class="_ _13"></span>ｉ<span class="_ _b"></span>ｏ<span class="_ _7"></span>ｎ</div><div class="t m0 xca he y4d ff3 fs5 fc2 sc2 ls0 ws0">　</div><div class="t m0 xb1 h11 y4a ff2 fs5 fc2 sc2 ls0 ws0">ｉ<span class="_ _13"></span>ｓ</div><div class="t m0 xb2 he y4d ff3 fs5 fc2 sc2 ls0 ws0">　</div><div class="t m0 x19 h11 y4e ff2 fs5 fc2 sc2 ls0 ws0">ｐ</div><div class="t m0 xcb h11 y4f ff2 fs5 fc2 sc2 ls0 ws0">ｒ<span class="_ _a"></span>ｅ<span class="_ _9"></span>ｓ<span class="_ _a"></span>ｅ<span class="_ _7"></span>ｎ<span class="_ _9"></span>ｔ<span class="_ _b"></span>ｅ<span class="_ _7"></span>ｄ</div><div class="t m0 xcc he y4d ff3 fs5 fc2 sc2 ls0 ws0">　</div><div class="t m0 xcd h11 y4a ff2 fs5 fc2 sc2 ls0 ws0">ｆ<span class="_ _b"></span>ｒ<span class="_ _a"></span>ｏ<span class="_ _6"></span>ｍ</div><div class="t m0 x1e he y4d ff3 fs5 fc2 sc2 ls0 ws0">　</div><div class="t m0 xce h11 y4a ff2 fs5 fc2 sc2 ls0 ws0">ｔ<span class="_ _9"></span>ｈ<span class="_ _7"></span>ｅ</div><div class="t m0 xcf he y4d ff3 fs5 fc2 sc2 ls0 ws0">　</div><div class="t m0 xd0 h11 y4a ff2 fs5 fc2 sc2 ls0 ws0">ｆ<span class="_ _a"></span>ｏ<span class="_ _b"></span>ｌ<span class="_ _12"></span>ｌ<span class="_ _b"></span>ｏ<span class="_ _4"></span>ｗ<span class="_ _8"></span>ｉ<span class="_ _b"></span>ｎ</div><div class="t m0 xd1 h11 y4b ff2 fs5 fc2 sc2 ls0 ws0">ｇ</div><div class="t m0 xd2 he y4c ff3 fs5 fc2 sc2 ls0 ws0">　</div><div class="t m0 xd3 h11 y4a ff2 fs5 fc2 sc2 ls0 ws0">ｆ<span class="_ _12"></span>ｉ<span class="_ _b"></span>ｖ<span class="_ _9"></span>ｅ</div><div class="t m0 xd4 he y4d ff3 fs5 fc2 sc2 ls0 ws0">　</div><div class="t m0 xd5 h11 y4e ff2 fs5 fc2 sc2 ls0 ws0">ｐ</div><div class="t m0 xd6 h11 y4f ff2 fs5 fc2 sc2 ls0 ws0">ｏ<span class="_ _b"></span>ｉ<span class="_ _b"></span>ｎ<span class="_ _a"></span>ｔ<span class="_ _b"></span>ｓ</div><div class="t m0 xd7 h16 y50 ff4 fs5 fc2 sc2 ls0 ws0">：</div><div class="t m0 xd8 h11 y4a ff2 fs5 fc2 sc2 ls0 ws0">Ｔ<span class="_ _1"></span>ｒ<span class="_ _a"></span>ａ<span class="_ _b"></span>ｉ<span class="_ _b"></span>ｎ<span class="_ _b"></span>ｉ<span class="_ _b"></span>ｎ</div><div class="t m0 x3f h11 y4b ff2 fs5 fc2 sc2 ls0 ws0">ｇ</div><div class="t m0 x36 he y4c ff3 fs5 fc2 sc2 ls0 ws0">　</div><div class="t m0 x37 h11 y4a ff2 fs5 fc2 sc2 ls0 ws0">ｃ<span class="_ _b"></span>ｒ<span class="_ _13"></span>ｉ<span class="_ _13"></span>ｔ<span class="_ _b"></span>ｅ<span class="_ _a"></span>ｒ<span class="_ _13"></span>ｉ<span class="_ _b"></span>ｏ<span class="_ _7"></span>ｎ<span class="_ _9"></span>ｓ</div><div class="t m0 xd9 he y4d ff3 fs5 fc2 sc2 ls0 ws0">　</div><div class="t m0 xda h11 y4a ff2 fs5 fc2 sc2 ls0 ws0">ｆ<span class="_ _a"></span>ｏ<span class="_ _a"></span>ｒ</div><div class="t m0 x2 h11 y51 ff2 fs5 fc2 sc2 ls0 ws0">ｄ<span class="_ _9"></span>ｅ<span class="_ _9"></span>ｅ</div><div class="t m0 xc2 h11 y52 ff2 fs5 fc2 sc2 ls0 ws0">ｐ</div><div class="t m0 x3 he y53 ff3 fs5 fc2 sc2 ls0 ws0">　</div><div class="t m0 xc3 h11 y51 ff2 fs5 fc2 sc2 ls0 ws0">ｌ<span class="_ _b"></span>ｅ<span class="_ _a"></span>ａ<span class="_ _a"></span>ｒ<span class="_ _7"></span>ｎ<span class="_ _b"></span>ｉ<span class="_ _b"></span>ｎ</div><div class="t m0 x23 h11 y52 ff2 fs5 fc2 sc2 ls0 ws0">ｇ</div><div class="t m0 xf he y53 ff3 fs5 fc2 sc2 ls0 ws0">　</div><div class="t m0 x24 h11 y51 ff2 fs5 fc2 sc2 ls0 ws0">ｂ<span class="_ _7"></span>ａ<span class="_ _9"></span>ｓ<span class="_ _a"></span>ｅ<span class="_ _7"></span>ｄ</div><div class="t m0 xc4 he y54 ff3 fs5 fc2 sc2 ls0 ws0">　</div><div class="t m0 x7c h11 y51 ff2 fs5 fc2 sc2 ls0 ws0">ａ<span class="_ _9"></span>ｃ<span class="_ _9"></span>ｏ<span class="_ _7"></span>ｕ<span class="_ _7"></span>ｓ<span class="_ _b"></span>ｔ<span class="_ _12"></span>ｉ<span class="_ _b"></span>ｃ</div><div class="t m0 xdb he y54 ff3 fs5 fc2 sc2 ls0 ws0">　</div><div class="t m0 xdc h11 y51 ff2 fs5 fc2 sc2 ls0 ws0">ｍ<span class="_ _6"></span>ｏ<span class="_ _7"></span>ｄ<span class="_ _7"></span>ｅ<span class="_ _b"></span>ｌ<span class="_ _13"></span>ｓ</div><div class="t m0 x8a h16 y55 ff4 fs5 fc2 sc2 ls0 ws0">，</div><div class="t m0 xdd h11 y51 ff2 fs5 fc2 sc2 ls0 ws0">ｄ<span class="_ _b"></span>ｉ<span class="_ _12"></span>ｆ<span class="_ _13"></span>ｆ<span class="_ _b"></span>ｅ<span class="_ _a"></span>ｒ<span class="_ _a"></span>ｅ<span class="_ _7"></span>ｎ<span class="_ _9"></span>ｔ</div><div class="t m0 x77 he y54 ff3 fs5 fc2 sc2 ls0 ws0">　</div><div class="t m0 xb3 h11 y51 ff2 fs5 fc2 sc2 ls0 ws0">ｍ<span class="_ _6"></span>ｏ<span class="_ _7"></span>ｄ<span class="_ _7"></span>ｅ<span class="_ _b"></span>ｌ</div><div class="t m0 xde he y54 ff3 fs5 fc2 sc2 ls0 ws0">　</div><div class="t m0 x5e h11 y51 ff2 fs5 fc2 sc2 ls0 ws0">ａ<span class="_ _a"></span>ｒ<span class="_ _a"></span>ｃ<span class="_ _7"></span>ｈ<span class="_ _b"></span>ｉ<span class="_ _13"></span>ｔ<span class="_ _b"></span>ｅ<span class="_ _a"></span>ｃ<span class="_ _b"></span>ｔ<span class="_ _9"></span>ｕ<span class="_ _9"></span>ｒ<span class="_ _a"></span>ｅ<span class="_ _9"></span>ｓ</div><div class="t m0 xdf he y54 ff3 fs5 fc2 sc2 ls0 ws0">　</div><div class="t m0 xe0 h11 y51 ff2 fs5 fc2 sc2 ls0 ws0">ｆ<span class="_ _a"></span>ｏ<span class="_ _a"></span>ｒ</div><div class="t m0 xb5 he y54 ff3 fs5 fc2 sc2 ls0 ws0">　</div><div class="t m0 xd2 h11 y51 ff2 fs5 fc2 sc2 ls0 ws0">ｄ<span class="_ _9"></span>ｅ<span class="_ _9"></span>ｅ</div><div class="t m0 xe1 h11 y52 ff2 fs5 fc2 sc2 ls0 ws0">ｐ</div><div class="t m0 xe2 he y53 ff3 fs5 fc2 sc2 ls0 ws0">　</div><div class="t m0 xe3 h11 y51 ff2 fs5 fc2 sc2 ls0 ws0">ｌ<span class="_ _b"></span>ｅ<span class="_ _a"></span>ａ<span class="_ _a"></span>ｒ<span class="_ _7"></span>ｎ<span class="_ _b"></span>ｉ<span class="_ _b"></span>ｎ</div><div class="t m0 xe4 h11 y52 ff2 fs5 fc2 sc2 ls0 ws0">ｇ</div><div class="t m0 xe5 he y53 ff3 fs5 fc2 sc2 ls0 ws0">　</div><div class="t m0 xbb h11 y51 ff2 fs5 fc2 sc2 ls0 ws0">ｂ<span class="_ _7"></span>ａ<span class="_ _9"></span>ｓ<span class="_ _a"></span>ｅ<span class="_ _7"></span>ｄ</div><div class="t m0 x35 he y54 ff3 fs5 fc2 sc2 ls0 ws0">　</div><div class="t m0 xe6 h11 y51 ff2 fs5 fc2 sc2 ls0 ws0">ｓ</div><div class="t m0 xe7 h11 y52 ff2 fs5 fc2 sc2 ls0 ws0">ｐ</div><div class="t m0 x37 h11 y51 ff2 fs5 fc2 sc2 ls0 ws0">ｅ<span class="_ _9"></span>ｅ<span class="_ _a"></span>ｃ<span class="_ _7"></span>ｈ</div><div class="t m0 xe8 he y54 ff3 fs5 fc2 sc2 ls0 ws0">　</div><div class="t m0 xe9 h11 y51 ff2 fs5 fc2 sc2 ls0 ws0">ｒ<span class="_ _a"></span>ｅ<span class="_ _9"></span>ｃ<span class="_ _9"></span>ｏ</div><div class="t m0 xea h11 y52 ff2 fs5 fc2 sc2 ls0 ws0">ｇ</div><div class="t m0 xc0 h11 y51 ff2 fs5 fc2 sc2 ls0 ws0">ｎ<span class="_ _b"></span>ｉ</div><div class="t m4 xeb h11 y56 ff2 fs5 fc2 sc2 ls0 ws0">－</div><div class="t m0 x2 h11 y57 ff2 fs5 fc2 sc2 ls0 ws0">ｔ<span class="_ _12"></span>ｉ<span class="_ _b"></span>ｏ<span class="_ _8"></span>ｎ</div><div class="t m0 x21 he y58 ff3 fs5 fc2 sc2 ls0 ws0">　</div><div class="t m0 xec h11 y57 ff2 fs5 fc2 sc2 ls0 ws0">ａ<span class="_ _9"></span>ｃ<span class="_ _9"></span>ｏ<span class="_ _7"></span>ｕ<span class="_ _7"></span>ｓ<span class="_ _b"></span>ｔ<span class="_ _12"></span>ｉ<span class="_ _b"></span>ｃ</div><div class="t m0 xed he y58 ff3 fs5 fc2 sc2 ls0 ws0">　</div><div class="t m0 x24 h11 y57 ff2 fs5 fc2 sc2 ls0 ws0">ｍ<span class="_ _6"></span>ｏ<span class="_ _7"></span>ｄ<span class="_ _7"></span>ｅ<span class="_ _b"></span>ｌ<span class="_ _15"></span>ｉ<span class="_ _a"></span>ｎ</div><div class="t m0 xa h11 y59 ff2 fs5 fc2 sc2 ls0 ws0">ｇ</div><div class="t m0 xee h16 y5a ff4 fs5 fc2 sc2 ls0 ws0">，</div><div class="t m0 x2d h11 y57 ff2 fs5 fc2 sc2 ls0 ws0">ｓ<span class="_ _a"></span>ｃ<span class="_ _9"></span>ａ<span class="_ _b"></span>ｌ<span class="_ _13"></span>ａ<span class="_ _7"></span>ｂ<span class="_ _a"></span>ｌ<span class="_ _b"></span>ｅ</div><div class="t m0 xef he y58 ff3 fs5 fc2 sc2 ls0 ws0">　</div><div class="t m0 xf0 h11 y57 ff2 fs5 fc2 sc2 ls0 ws0">ａ<span class="_ _7"></span>ｎ<span class="_ _8"></span>ｄ</div><div class="t m0 x8b he y58 ff3 fs5 fc2 sc2 ls0 ws0">　</div><div class="t m0 xb0 h11 y57 ff2 fs5 fc2 sc2 ls0 ws0">ｄ<span class="_ _b"></span>ｉ<span class="_ _13"></span>ｓ<span class="_ _b"></span>ｔ<span class="_ _b"></span>ｒ<span class="_ _13"></span>ｉ<span class="_ _b"></span>ｂ<span class="_ _8"></span>ｕ<span class="_ _a"></span>ｔ<span class="_ _a"></span>ｅ<span class="_ _7"></span>ｄ</div><div class="t m0 xf1 he y58 ff3 fs5 fc2 sc2 ls0 ws0">　</div><div class="t m0 xf2 h11 y57 ff2 fs5 fc2 sc2 ls0 ws0">ｏ</div><div class="t m0 xcc h11 y59 ff2 fs5 fc2 sc2 ls0 ws0">ｐ</div><div class="t m0 xf3 h11 y57 ff2 fs5 fc2 sc2 ls0 ws0">ｔ<span class="_ _12"></span>ｉ<span class="_ _1"></span>ｍ<span class="_ _1"></span>ｉ<span class="_ _13"></span>ｚ<span class="_ _a"></span>ａ<span class="_ _a"></span>ｔ<span class="_ _12"></span>ｉ<span class="_ _b"></span>ｏ<span class="_ _8"></span>ｎ</div><div class="t m0 xf4 he y58 ff3 fs5 fc2 sc2 ls0 ws0">　</div><div class="t m0 xf5 h11 y57 ff2 fs5 fc2 sc2 ls0 ws0">ｍ<span class="_ _2"></span>ｅ<span class="_ _a"></span>ｔ<span class="_ _9"></span>ｈ<span class="_ _7"></span>ｏ<span class="_ _8"></span>ｄ<span class="_ _9"></span>ｓ</div><div class="t m0 xf6 he y58 ff3 fs5 fc2 sc2 ls0 ws0">　</div><div class="t m0 xf7 h11 y57 ff2 fs5 fc2 sc2 ls0 ws0">ｆ<span class="_ _a"></span>ｏ<span class="_ _a"></span>ｒ</div><div class="t m0 xd5 he y58 ff3 fs5 fc2 sc2 ls0 ws0">　</div><div class="t m0 xd6 h11 y57 ff2 fs5 fc2 sc2 ls0 ws0">ｄ<span class="_ _9"></span>ｅ<span class="_ _9"></span>ｅ</div><div class="t m0 xf8 h11 y59 ff2 fs5 fc2 sc2 ls0 ws0">ｐ</div><div class="t m0 xba he y5b ff3 fs5 fc2 sc2 ls0 ws0">　</div><div class="t m0 xf9 h11 y57 ff2 fs5 fc2 sc2 ls0 ws0">ｌ<span class="_ _b"></span>ｅ<span class="_ _a"></span>ａ<span class="_ _a"></span>ｒ<span class="_ _7"></span>ｎ<span class="_ _b"></span>ｉ<span class="_ _b"></span>ｎ</div><div class="t m0 xfa h11 y59 ff2 fs5 fc2 sc2 ls0 ws0">ｇ</div><div class="t m0 xe6 he y5b ff3 fs5 fc2 sc2 ls0 ws0">　</div><div class="t m0 xfb h11 y57 ff2 fs5 fc2 sc2 ls0 ws0">ｂ<span class="_ _7"></span>ａ<span class="_ _9"></span>ｓ<span class="_ _a"></span>ｅ<span class="_ _7"></span>ｄ</div><div class="t m0 xe8 he y58 ff3 fs5 fc2 sc2 ls0 ws0">　</div><div class="t m0 x49 h11 y57 ff2 fs5 fc2 sc2 ls0 ws0">ａ<span class="_ _9"></span>ｃ<span class="_ _9"></span>ｏ<span class="_ _7"></span>ｕ<span class="_ _7"></span>ｓ<span class="_ _b"></span>ｔ<span class="_ _12"></span>ｉ<span class="_ _b"></span>ｃ</div><div class="t m0 x20 h11 y5c ff2 fs5 fc2 sc2 ls0 ws0">ｍ<span class="_ _6"></span>ｏ<span class="_ _7"></span>ｄ<span class="_ _7"></span>ｅ<span class="_ _b"></span>ｌ</div><div class="t m0 xb he y5d ff3 fs5 fc2 sc2 ls0 ws0">　</div><div class="t m0 xc h11 y5e ff2 fs5 fc2 sc2 ls0 ws0">ｔ<span class="_ _b"></span>ｒ<span class="_ _a"></span>ａ<span class="_ _13"></span>ｉ<span class="_ _a"></span>ｎ<span class="_ _b"></span>ｉ<span class="_ _b"></span>ｎ</div><div class="t m0 xf h11 y5f ff2 fs5 fc2 sc2 ls0 ws0">ｇ</div><div class="t m0 x7b h16 y60 ff4 fs5 fc2 sc2 ls0 ws0">，</div><div class="t m0 x11 h11 y5e ff2 fs5 fc2 sc2 ls0 ws0">ｓ</div><div class="t m0 xa2 h11 y5f ff2 fs5 fc2 sc2 ls0 ws0">ｐ</div><div class="t m0 x25 h11 y5e ff2 fs5 fc2 sc2 ls0 ws0">ｅ<span class="_ _9"></span>ａ<span class="_ _7"></span>ｋ<span class="_ _7"></span>ｅ<span class="_ _a"></span>ｒ</div><div class="t m0 x6b he y5d ff3 fs5 fc2 sc2 ls0 ws0">　</div><div class="t m0 xfc h11 y5e ff2 fs5 fc2 sc2 ls0 ws0">ａ<span class="_ _7"></span>ｄ<span class="_ _9"></span>ａ</div><div class="t m0 x88 h11 y5f ff2 fs5 fc2 sc2 ls0 ws0">ｐ</div><div class="t m0 xfd h11 y5e ff2 fs5 fc2 sc2 ls0 ws0">ｔ<span class="_ _b"></span>ａ<span class="_ _a"></span>ｔ<span class="_ _13"></span>ｉ<span class="_ _13"></span>ｏ<span class="_ _8"></span>ｎ</div><div class="t m0 xfe he y5d ff3 fs5 fc2 sc2 ls0 ws0">　</div><div class="t m0 xff h11 y5e ff2 fs5 fc2 sc2 ls0 ws0">ｆ<span class="_ _a"></span>ｏ<span class="_ _a"></span>ｒ</div><div class="t m0 xca he y5d ff3 fs5 fc2 sc2 ls0 ws0">　</div><div class="t m0 xb1 h11 y5e ff2 fs5 fc2 sc2 ls0 ws0">ｄ<span class="_ _9"></span>ｅ<span class="_ _9"></span>ｅ</div><div class="t m0 x1a h11 y5f ff2 fs5 fc2 sc2 ls0 ws0">ｐ</div><div class="t m0 x100 he y61 ff3 fs5 fc2 sc2 ls0 ws0">　</div><div class="t m0 x101 h11 y5e ff2 fs5 fc2 sc2 ls0 ws0">ｌ<span class="_ _b"></span>ｅ<span class="_ _a"></span>ａ<span class="_ _a"></span>ｒ<span class="_ _7"></span>ｎ<span class="_ _b"></span>ｉ<span class="_ _b"></span>ｎ</div><div class="t m0 xcc h11 y5f ff2 fs5 fc2 sc2 ls0 ws0">ｇ</div><div class="t m0 x102 he y61 ff3 fs5 fc2 sc2 ls0 ws0">　</div><div class="t m0 x103 h11 y5e ff2 fs5 fc2 sc2 ls0 ws0">ｂ<span class="_ _7"></span>ａ<span class="_ _9"></span>ｓ<span class="_ _a"></span>ｅ<span class="_ _7"></span>ｄ</div><div class="t m0 x53 he y5d ff3 fs5 fc2 sc2 ls0 ws0">　</div><div class="t m0 x54 h11 y5e ff2 fs5 fc2 sc2 ls0 ws0">ａ<span class="_ _9"></span>ｃ<span class="_ _9"></span>ｏ<span class="_ _7"></span>ｕ<span class="_ _7"></span>ｓ<span class="_ _b"></span>ｔ<span class="_ _12"></span>ｉ<span class="_ _b"></span>ｃ</div><div class="t m0 x5b he y5d ff3 fs5 fc2 sc2 ls0 ws0">　</div><div class="t m0 x104 h11 y5e ff2 fs5 fc2 sc2 ls0 ws0">ｍ<span class="_ _6"></span>ｏ<span class="_ _7"></span>ｄ<span class="_ _7"></span>ｅ<span class="_ _b"></span>ｌ</div><div class="t m0 x105 h16 y60 ff4 fs5 fc2 sc2 ls0 ws0">，</div><div class="t m0 xe3 h11 y5e ff2 fs5 fc2 sc2 ls0 ws0">ａ<span class="_ _7"></span>ｎ<span class="_ _8"></span>ｄ</div><div class="t m0 x106 he y5d ff3 fs5 fc2 sc2 ls0 ws0">　</div><div class="t m0 x107 h11 y5e ff2 fs5 fc2 sc2 ls0 ws0">ｄ<span class="_ _9"></span>ｅ<span class="_ _9"></span>ｅ</div><div class="t m0 x108 h11 y5f ff2 fs5 fc2 sc2 ls0 ws0">ｐ</div><div class="t m0 x109 he y61 ff3 fs5 fc2 sc2 ls0 ws0">　</div><div class="t m0 x10a h11 y5e ff2 fs5 fc2 sc2 ls0 ws0">ｌ<span class="_ _b"></span>ｅ<span class="_ _a"></span>ａ<span class="_ _7"></span>ｎ<span class="_ _b"></span>ｉ<span class="_ _a"></span>ｎ</div><div class="t m0 x10b h11 y5f ff2 fs5 fc2 sc2 ls0 ws0">ｇ</div><div class="t m0 x36 he y61 ff3 fs5 fc2 sc2 ls0 ws0">　</div><div class="t m0 x37 h11 y5e ff2 fs5 fc2 sc2 ls0 ws0">ｂ<span class="_ _7"></span>ａ<span class="_ _9"></span>ｓ<span class="_ _a"></span>ｅ<span class="_ _7"></span>ｄ</div><div class="t m0 x10c he y5d ff3 fs5 fc2 sc2 ls0 ws0">　</div><div class="t m0 x10d h11 y5e ff2 fs5 fc2 sc2 ls0 ws0">ｅ<span class="_ _7"></span>ｎ<span class="_ _8"></span>ｄ</div><div class="t m4 x10e h11 y62 ff2 fs5 fc2 sc2 ls0 ws0">－</div><div class="t m0 xda h11 y5e ff2 fs5 fc2 sc2 ls0 ws0">ｔ<span class="_ _a"></span>ｏ</div><div class="t m4 xeb h11 y62 ff2 fs5 fc2 sc2 ls0 ws0">－</div><div class="t m0 x2 h11 y63 ff2 fs5 fc2 sc2 ls0 ws0">ｅ<span class="_ _7"></span>ｎ<span class="_ _8"></span>ｄ</div><div class="t m0 xc2 he y64 ff3 fs5 fc2 sc2 ls0 ws0">　</div><div class="t m0 x22 h11 y63 ff2 fs5 fc2 sc2 ls0 ws0">ｓ</div><div class="t m0 x10f h11 y65 ff2 fs5 fc2 sc2 ls0 ws0">ｐ</div><div class="t m0 xc h11 y63 ff2 fs5 fc2 sc2 ls0 ws0">ｅ<span class="_ _9"></span>ｅ<span class="_ _a"></span>ｃ<span class="_ _7"></span>ｈ</div><div class="t m0 x110 he y64 ff3 fs5 fc2 sc2 ls0 ws0">　</div><div class="t m0 x111 h11 y63 ff2 fs5 fc2 sc2 ls0 ws0">ｒ<span class="_ _a"></span>ｅ<span class="_ _9"></span>ｃ<span class="_ _9"></span>ｏ</div><div class="t m0 x11 h11 y65 ff2 fs5 fc2 sc2 ls0 ws0">ｇ</div><div class="t m0 xa2 h11 y63 ff2 fs5 fc2 sc2 ls0 ws0">ｎ<span class="_ _b"></span>ｉ<span class="_ _12"></span>ｔ<span class="_ _13"></span>ｉ<span class="_ _b"></span>ｏ<span class="_ _7"></span>ｎ<span class="_ _3"></span>．<span class="_ _16"></span>Ａ<span class="_ _1"></span>ｔ</div><div class="t m0 x92 he y64 ff3 fs5 fc2 sc2 ls0 ws0">　</div><div class="t m0 xa5 h11 y63 ff2 fs5 fc2 sc2 ls0 ws0">ｔ<span class="_ _9"></span>ｈ<span class="_ _7"></span>ｅ</div><div class="t m0 x74 he y64 ff3 fs5 fc2 sc2 ls0 ws0">　</div><div class="t m0 xc9 h11 y63 ff2 fs5 fc2 sc2 ls0 ws0">ｅ<span class="_ _7"></span>ｎ<span class="_ _8"></span>ｄ</div><div class="t m0 x8a he y64 ff3 fs5 fc2 sc2 ls0 ws0">　</div><div class="t m0 x75 h11 y63 ff2 fs5 fc2 sc2 ls0 ws0">ｏ<span class="_ _b"></span>ｆ</div><div class="t m0 x112 he y64 ff3 fs5 fc2 sc2 ls0 ws0">　</div><div class="t m0 x97 h11 y63 ff2 fs5 fc2 sc2 ls0 ws0">ｔ<span class="_ _9"></span>ｈ<span class="_ _b"></span>ｉ<span class="_ _13"></span>ｓ</div><div class="t m0 x6d he y64 ff3 fs5 fc2 sc2 ls0 ws0">　</div><div class="t m0 x113 h11 y66 ff2 fs5 fc2 sc2 ls0 ws0">ｐ</div><div class="t m0 x114 h11 y67 ff2 fs5 fc2 sc2 ls0 ws0">ａ</div><div class="t m0 x82 h11 y66 ff2 fs5 fc2 sc2 ls0 ws0">ｐ</div><div class="t m0 x78 h11 y67 ff2 fs5 fc2 sc2 ls0 ws0">ｅ<span class="_ _a"></span>ｒ</div><div class="t m0 xcc h16 y68 ff4 fs5 fc2 sc2 ls0 ws0">，</div><div class="t m0 xde h11 y63 ff2 fs5 fc2 sc2 ls0 ws0">ｔ<span class="_ _9"></span>ｈ<span class="_ _7"></span>ｅ</div><div class="t m0 x64 he y64 ff3 fs5 fc2 sc2 ls0 ws0">　</div><div class="t m0 x115 h11 y63 ff2 fs5 fc2 sc2 ls0 ws0">ｆ<span class="_ _a"></span>ｕ<span class="_ _9"></span>ｔ<span class="_ _9"></span>ｕ<span class="_ _9"></span>ｒ<span class="_ _a"></span>ｅ</div><div class="t m0 x116 he y64 ff3 fs5 fc2 sc2 ls0 ws0">　</div><div class="t m0 x117 h11 y66 ff2 fs5 fc2 sc2 ls0 ws0">ｐ</div><div class="t m0 x5f h11 y67 ff2 fs5 fc2 sc2 ls0 ws0">ｏ<span class="_ _9"></span>ｓ<span class="_ _a"></span>ｓ<span class="_ _13"></span>ｉ<span class="_ _b"></span>ｂ<span class="_ _a"></span>ｌ<span class="_ _b"></span>ｅ</div><div class="t m0 x118 he y64 ff3 fs5 fc2 sc2 ls0 ws0">　</div><div class="t m0 x119 h11 y63 ff2 fs5 fc2 sc2 ls0 ws0">ｒ<span class="_ _a"></span>ｅ<span class="_ _9"></span>ｓ<span class="_ _a"></span>ｅ<span class="_ _9"></span>ａ<span class="_ _a"></span>ｒ<span class="_ _a"></span>ｃ<span class="_ _7"></span>ｈ</div><div class="t m0 x11a he y64 ff3 fs5 fc2 sc2 ls0 ws0">　</div><div class="t m0 xe4 h11 y66 ff2 fs5 fc2 sc2 ls0 ws0">ｐ</div><div class="t m0 x11b h11 y67 ff2 fs5 fc2 sc2 ls0 ws0">ｏ<span class="_ _b"></span>ｉ<span class="_ _b"></span>ｎ<span class="_ _a"></span>ｔ<span class="_ _b"></span>ｓ</div><div class="t m0 x34 he y64 ff3 fs5 fc2 sc2 ls0 ws0">　</div><div class="t m0 x11c h11 y63 ff2 fs5 fc2 sc2 ls0 ws0">ｏ<span class="_ _b"></span>ｆ</div><div class="t m0 xfb he y64 ff3 fs5 fc2 sc2 ls0 ws0">　</div><div class="t m0 x41 h11 y63 ff2 fs5 fc2 sc2 ls0 ws0">ｄ<span class="_ _9"></span>ｅ<span class="_ _9"></span>ｅ</div><div class="t m0 x44 h11 y65 ff2 fs5 fc2 sc2 ls0 ws0">ｐ</div><div class="t m0 x11d he y69 ff3 fs5 fc2 sc2 ls0 ws0">　</div><div class="t m0 xe9 h11 y63 ff2 fs5 fc2 sc2 ls0 ws0">ｌ<span class="_ _b"></span>ｅ<span class="_ _a"></span>ａ<span class="_ _a"></span>ｒ<span class="_ _7"></span>ｎ<span class="_ _b"></span>ｉ<span class="_ _b"></span>ｎ</div><div class="t m0 x11e h11 y65 ff2 fs5 fc2 sc2 ls0 ws0">ｇ</div><div class="t m0 x2 h11 y6a ff2 fs5 fc2 sc2 ls0 ws0">ｂ<span class="_ _7"></span>ａ<span class="_ _9"></span>ｓ<span class="_ _a"></span>ｅ<span class="_ _7"></span>ｄ</div><div class="t m0 x10f he y6b ff3 fs5 fc2 sc2 ls0 ws0">　</div><div class="t m0 x11f h11 y6a ff2 fs5 fc2 sc2 ls0 ws0">ｓ</div><div class="t m0 x120 h11 y6c ff2 fs5 fc2 sc2 ls0 ws0">ｐ</div><div class="t m0 x8e h11 y6a ff2 fs5 fc2 sc2 ls0 ws0">ｅ<span class="_ _9"></span>ｅ<span class="_ _a"></span>ｃ<span class="_ _7"></span>ｈ</div><div class="t m0 xe he y6b ff3 fs5 fc2 sc2 ls0 ws0">　</div><div class="t m0 x7 h11 y6a ff2 fs5 fc2 sc2 ls0 ws0">ｒ<span class="_ _a"></span>ｅ<span class="_ _9"></span>ｃ<span class="_ _9"></span>ｏ</div><div class="t m0 x2b h11 y6c ff2 fs5 fc2 sc2 ls0 ws0">ｇ</div><div class="t m0 x26 h11 y6a ff2 fs5 fc2 sc2 ls0 ws0">ｎ<span class="_ _b"></span>ｉ<span class="_ _12"></span>ｔ<span class="_ _13"></span>ｉ<span class="_ _b"></span>ｏ<span class="_ _7"></span>ｎ</div><div class="t m0 xfc he y6b ff3 fs5 fc2 sc2 ls0 ws0">　</div><div class="t m0 x121 h11 y6a ff2 fs5 fc2 sc2 ls0 ws0">ａ<span class="_ _a"></span>ｒ<span class="_ _a"></span>ｅ</div><div class="t m0 xdb he y6b ff3 fs5 fc2 sc2 ls0 ws0">　</div><div class="t m0 x66 h11 y6a ff2 fs5 fc2 sc2 ls0 ws0">ａ<span class="_ _b"></span>ｌ<span class="_ _13"></span>ｓ<span class="_ _9"></span>ｏ</div><div class="t m0 x122 he y6b ff3 fs5 fc2 sc2 ls0 ws0">　</div><div class="t m0 x123 h11 y6d ff2 fs5 fc2 sc2 ls0 ws0">ｐ</div><div class="t m0 x17 h11 y6e ff2 fs5 fc2 sc2 ls0 ws0">ｒ<span class="_ _9"></span>ｏ</div><div class="t m0 x75 h11 y6d ff2 fs5 fc2 sc2 ls0 ws0">ｐ</div><div class="t m0 x124 h11 y6e ff2 fs5 fc2 sc2 ls0 ws0">ｏ<span class="_ _9"></span>ｓ<span class="_ _a"></span>ｅ<span class="_ _7"></span>ｄ<span class="_ _3"></span>．</div><div class="t m0 x20 h17 y6f ffc fs5 fc2 sc2 ls0 ws0">Ｋ<span class="_ _1"></span>ｅ</div><div class="t m0 x125 h17 y70 ffc fs5 fc2 sc2 ls0 ws0">ｙ</div><div class="t m0 x21 he y71 ff3 fs5 fc2 sc2 ls0 ws0">　</div><div class="t m0 x4 h17 y6f ffc fs5 fc2 sc2 ls0 ws0">ｗ<span class="_ _1"></span>ｏ<span class="_ _9"></span>ｒ<span class="_ _a"></span>ｄ<span class="_ _9"></span>ｓ</div><div class="t m0 x126 h16 y72 ff4 fs5 fc2 sc2 ls0 ws0">：</div><div class="t m0 x127 h11 y6f ff2 fs5 fc2 sc2 ls0 ws0">ｄ<span class="_ _9"></span>ｅ<span class="_ _9"></span>ｅ</div><div class="t m0 x128 h11 y73 ff2 fs5 fc2 sc2 ls0 ws0">ｐ</div><div class="t m0 x129 he y74 ff3 fs5 fc2 sc2 ls0 ws0">　</div><div class="t m0 x70 h11 y6f ff2 fs5 fc2 sc2 ls0 ws0">ｌ<span class="_ _b"></span>ｅ<span class="_ _a"></span>ａ<span class="_ _a"></span>ｒ<span class="_ _7"></span>ｎ<span class="_ _b"></span>ｉ<span class="_ _b"></span>ｎ</div><div class="t m0 x72 h11 y73 ff2 fs5 fc2 sc2 ls0 ws0">ｇ</div><div class="t m0 x12a h16 y72 ff4 fs5 fc2 sc2 ls0 ws0">；</div><div class="t m0 x92 h11 y6f ff2 fs5 fc2 sc2 ls0 ws0">ｄ<span class="_ _9"></span>ｅ<span class="_ _9"></span>ｅ</div><div class="t m0 x12b h11 y73 ff2 fs5 fc2 sc2 ls0 ws0">ｐ</div><div class="t m0 x74 he y74 ff3 fs5 fc2 sc2 ls0 ws0">　</div><div class="t m0 x12c h11 y6f ff2 fs5 fc2 sc2 ls0 ws0">ｎ<span class="_ _7"></span>ｅ<span class="_ _7"></span>ｕ<span class="_ _9"></span>ｒ<span class="_ _a"></span>ａ<span class="_ _b"></span>ｌ</div><div class="t m0 xb1 he y75 ff3 fs5 fc2 sc2 ls0 ws0">　</div><div class="t m0 x12d h11 y6f ff2 fs5 fc2 sc2 ls0 ws0">ｎ<span class="_ _7"></span>ｅ<span class="_ _b"></span>ｔ<span class="_ _1"></span>ｗ<span class="_ _2"></span>ｏ<span class="_ _a"></span>ｒ<span class="_ _7"></span>ｋ</div><div class="t m0 x78 h16 y72 ff4 fs5 fc2 sc2 ls0 ws0">；</div><div class="t m0 xf2 h11 y6f ff2 fs5 fc2 sc2 ls0 ws0">ｓ</div><div class="t m0 x5d h11 y73 ff2 fs5 fc2 sc2 ls0 ws0">ｐ</div><div class="t m0 xf3 h11 y6f ff2 fs5 fc2 sc2 ls0 ws0">ｅ<span class="_ _9"></span>ｅ<span class="_ _a"></span>ｃ<span class="_ _7"></span>ｈ</div><div class="t m0 x1e he y75 ff3 fs5 fc2 sc2 ls0 ws0">　</div><div class="t m0 xce h11 y6f ff2 fs5 fc2 sc2 ls0 ws0">ｒ<span class="_ _a"></span>ｅ<span class="_ _9"></span>ｃ<span class="_ _9"></span>ｏ</div><div class="t m0 x12e h11 y73 ff2 fs5 fc2 sc2 ls0 ws0">ｇ</div><div class="t m0 xf4 h11 y6f ff2 fs5 fc2 sc2 ls0 ws0">ｎ<span class="_ _b"></span>ｉ<span class="_ _12"></span>ｔ<span class="_ _13"></span>ｉ<span class="_ _b"></span>ｏ<span class="_ _7"></span>ｎ</div><div class="t m0 x9c h16 y72 ff4 fs5 fc2 sc2 ls0 ws0">；</div><div class="t m0 x12f h11 y6f ff2 fs5 fc2 sc2 ls0 ws0">ｓ</div><div class="t m0 x130 h11 y73 ff2 fs5 fc2 sc2 ls0 ws0">ｐ</div><div class="t m0 xf7 h11 y6f ff2 fs5 fc2 sc2 ls0 ws0">ｅ<span class="_ _9"></span>ａ<span class="_ _7"></span>ｋ<span class="_ _7"></span>ｅ<span class="_ _a"></span>ｒ</div><div class="t m0 x131 he y75 ff3 fs5 fc2 sc2 ls0 ws0">　</div><div class="t m0 x132 h11 y6f ff2 fs5 fc2 sc2 ls0 ws0">ａ<span class="_ _7"></span>ｄ<span class="_ _9"></span>ａ</div><div class="t m0 xf9 h11 y73 ff2 fs5 fc2 sc2 ls0 ws0">ｐ</div><div class="t m0 x108 h11 y6f ff2 fs5 fc2 sc2 ls0 ws0">ｔ<span class="_ _b"></span>ａ<span class="_ _a"></span>ｔ<span class="_ _13"></span>ｉ<span class="_ _13"></span>ｏ<span class="_ _8"></span>ｎ</div><div class="t m0 x27 h18 y76 ff6 fs6 fc2 sc2 ls0 ws0">引</div><div class="t m0 xc2 h13 y77 ff3 fs6 fc2 sc2 ls0 ws0">　<span class="_ _f"></span>　</div><div class="t m0 x6 h18 y76 ff6 fs6 fc2 sc2 ls0 ws0">言</div><div class="t m0 x4 h11 y78 ff2 fs5 fc2 sc2 ls0 ws0">２<span class="_ _7"></span>０<span class="_ _8"></span>０<span class="_ _8"></span>６</div><div class="t m0 x4d hd y79 ff8 fs5 fc2 sc2 ls0 ws0">年</div><div class="t m0 x133 h16 y7a ff4 fs5 fc2 sc2 ls0 ws0">，</div><div class="t m0 x8 hd y79 ff8 fs5 fc2 sc2 ls0 ws0">由<span class="_ _f"></span>于<span class="_ _f"></span>深<span class="_ _e"></span>度<span class="_ _f"></span>学<span class="_ _f"></span>习</div><div class="t m0 xa5 h19 y7b ff4 fs7 fc2 sc2 ls0 ws0">［</div><div class="t m0 xfd h1a y7c ff2 fs7 fc2 sc2 ls0 ws0">１</div><div class="t m5 x134 h1a y7d ff2 fs7 fc2 sc2 ls0 ws0">－</div><div class="t m0 xaf h1a y7c ff2 fs7 fc2 sc2 ls0 ws0">２</div><div class="t m0 x135 h19 y7b ff4 fs7 fc2 sc2 ls0 ws0">］</div><div class="t m0 x74 hd y79 ff8 fs5 fc2 sc2 ls4 ws0">理论在机器学习中初步的<span class="_ _10"></span>成功<span class="_ _10"></span>应用</div><div class="t m0 x136 h16 y7a ff4 fs5 fc2 sc2 ls0 ws0">，</div><div class="t m0 xd1 hd y79 ff8 fs5 fc2 sc2 ls0 ws0">开<span class="_ _e"></span>始<span class="_ _e"></span>引<span class="_ _e"></span>起<span class="_ _e"></span>人<span class="_ _17"></span>们<span class="_ _e"></span>的<span class="_ _e"></span>关<span class="_ _17"></span>注</div><div class="t m0 x40 h16 y7a ff4 fs5 fc2 sc2 ls0 ws0">。</div><div class="t m0 x39 hd y79 ff8 fs5 fc2 sc2 ls0 ws0">在<span class="_ _e"></span>接<span class="_ _e"></span>下<span class="_ _17"></span>来<span class="_ _e"></span>的</div><div class="t m0 x27 hd y7e ff8 fs5 fc2 sc2 ls0 ws0">几<span class="_ _f"></span>年<span class="_ _f"></span>里</div><div class="t m0 x58 h16 y7f ff4 fs5 fc2 sc2 ls0 ws0">，</div><div class="t m0 x5 hd y7e ff8 fs5 fc2 sc2 ls4 ws0">机器学习领域的研究热点开始逐步地转向深<span class="_ _10"></span>度学<span class="_ _10"></span>习</div><div class="t m0 x137 h16 y7f ff4 fs5 fc2 sc2 ls0 ws0">。</div><div class="t m0 x138 hd y7e ff8 fs5 fc2 sc2 ls6 ws0">深度学习使用多层的非线性结构将低层</div><div class="t m0 x27 hd y80 ff8 fs5 fc2 sc2 ls4 ws0">特征变换成更加抽象的高层特征</div><div class="t m0 x123 h16 y81 ff4 fs5 fc2 sc2 ls0 ws0">，</div><div class="t m0 xff hd y80 ff8 fs5 fc2 sc2 ls4 ws0">以有监督或者无监督的方法对输入特征进行变换</div><div class="t m0 x61 h16 y81 ff4 fs5 fc2 sc2 ls0 ws0">，</div><div class="t m0 x3d hd y80 ff8 fs5 fc2 sc2 ls0 ws0">从<span class="_ _f"></span>而<span class="_ _f"></span>提<span class="_ _e"></span>升<span class="_ _f"></span>分<span class="_ _e"></span>类<span class="_ _e"></span>或<span class="_ _e"></span>者</div><div class="t m0 x27 hd y82 ff8 fs5 fc2 sc2 ls0 ws0">预<span class="_ _f"></span>测<span class="_ _f"></span>的<span class="_ _e"></span>准<span class="_ _f"></span>确<span class="_ _f"></span>性</div><div class="t m0 x7b h19 y83 ff4 fs7 fc2 sc2 ls0 ws0">［</div><div class="t m0 x10 h1a y84 ff2 fs7 fc2 sc2 ls0 ws0">３</div><div class="t m0 x11 h19 y83 ff4 fs7 fc2 sc2 ls0 ws0">］</div><div class="t m0 x70 h16 y85 ff4 fs5 fc2 sc2 ls0 ws0">。</div><div class="t m0 xc4 hd y82 ff8 fs5 fc2 sc2 ls4 ws0">深度学习模型一般是指更深<span class="_ _10"></span>层<span class="_ _10"></span>的结<span class="_ _10"></span>构</div><div class="t m0 x139 h19 y83 ff4 fs7 fc2 sc2 ls0 ws0">［</div><div class="t m0 x13a h1a y84 ff2 fs7 fc2 sc2 ls0 ws0">４</div><div class="t m0 x13b h19 y83 ff4 fs7 fc2 sc2 ls0 ws0">］</div><div class="t m0 x13c hd y82 ff8 fs5 fc2 sc2 ls0 ws0">模<span class="_ _e"></span>型</div><div class="t m0 x13d h16 y85 ff4 fs5 fc2 sc2 ls0 ws0">，</div><div class="t m0 xe0 hd y82 ff8 fs5 fc2 sc2 ls6 ws0">它比传统的浅层模型拥有更多层的非</div><div class="t m0 x27 hd y86 ff8 fs5 fc2 sc2 ls0 ws0">线<span class="_ _f"></span>性<span class="_ _f"></span>变<span class="_ _e"></span>换</div><div class="t m0 xa1 h16 y87 ff4 fs5 fc2 sc2 ls0 ws0">，</div><div class="t m0 x86 hd y86 ff8 fs5 fc2 sc2 ls4 ws0">在表达和建模能力上更加强大</div><div class="t m0 x13e h16 y87 ff4 fs5 fc2 sc2 ls0 ws0">，</div><div class="t m0 x13f hd y86 ff8 fs5 fc2 sc2 ls4 ws0">在复杂信号的处理上会更具优势</div><div class="t m0 xb7 h16 y87 ff4 fs5 fc2 sc2 ls0 ws0">。</div><div class="t m0 x3 h1b y88 ff6 fs1 fc2 sc2 ls0 ws0">基<span class="_ _f"></span>金<span class="_ _10"></span>项<span class="_ _f"></span>目</div><div class="t m0 x111 h1c y89 ff9 fs1 fc2 sc2 ls0 ws0">：</div><div class="t m0 xed h1b y88 ff8 fs1 fc2 sc2 ls7 ws0">安徽省科技重大专项</div><div class="t m0 xaf h5 y89 ff4 fs1 fc2 sc2 ls0 ws0">（</div><div class="t m0 x140 h3 y8a ff2 fs1 fc2 sc2 ls0 ws0">１<span class="_ _4"></span>５<span class="_ _4"></span>ｃ<span class="_ _7"></span>ｚ<span class="_ _8"></span>ｚ<span class="_ _1"></span>０<span class="_ _4"></span>２<span class="_ _4"></span>０<span class="_ _4"></span>０<span class="_ _4"></span>７</div><div class="t m0 x97 h5 y89 ff4 fs1 fc2 sc2 ls0 ws0">）</div><div class="t m0 x141 h1b y88 ff8 fs1 fc2 sc2 ls0 ws0">资<span class="_ _f"></span>助<span class="_ _10"></span>项<span class="_ _f"></span>目</div><div class="t m0 x83 h5 y89 ff4 fs1 fc2 sc2 ls0 ws0">；</div><div class="t m0 x142 h1b y88 ff8 fs1 fc2 sc2 ls7 ws0">国家重点研发计划</div><div class="t m0 x5a h5 y89 ff4 fs1 fc2 sc2 ls0 ws0">（</div><div class="t m0 x13d h3 y8a ff2 fs1 fc2 sc2 ls1 ws0">２０１６<span class="_ _e"> </span>Ｙ<span class="_ _17"> </span>Ｆ<span class="_ _f"> </span>Ｂ<span class="_ _10"></span>１００１３００</div><div class="t m0 xb9 h5 y89 ff4 fs1 fc2 sc2 ls0 ws0">）</div><div class="t m0 xba h1b y88 ff8 fs1 fc2 sc2 ls0 ws0">资<span class="_ _f"></span>助<span class="_ _10"></span>项<span class="_ _f"></span>目</div><div class="t m0 x143 h5 y89 ff4 fs1 fc2 sc2 ls0 ws0">。</div><div class="t m0 x3 h1b y8b ff6 fs1 fc2 sc2 ls0 ws0">收<span class="_ _f"></span>稿<span class="_ _10"></span>日<span class="_ _f"></span>期</div><div class="t m0 x111 h1c y8c ff9 fs1 fc2 sc2 ls0 ws0">：</div><div class="t m0 x62 h3 y8d ff2 fs1 fc2 sc2 ls0 ws0">２<span class="_ _4"></span>０<span class="_ _4"></span>１<span class="_ _4"></span>７</div><div class="t m1 x129 h3 y8e ff2 fs1 fc2 sc2 ls0 ws0">－</div><div class="t m0 x144 h3 y8d ff2 fs1 fc2 sc2 ls0 ws0">０<span class="_ _4"></span>１</div><div class="t m1 x26 h3 y8e ff2 fs1 fc2 sc2 ls0 ws0">－</div><div class="t m0 xc4 h3 y8d ff2 fs1 fc2 sc2 ls0 ws0">０<span class="_ _4"></span>５</div><div class="t m0 x71 h5 y8c ff4 fs1 fc2 sc2 ls0 ws0">；</div><div class="t m0 x13 h1b y8b ff6 fs1 fc2 sc2 ls0 ws0">修<span class="_ _f"></span>订<span class="_ _10"></span>日<span class="_ _f"></span>期</div><div class="t m0 x94 h1c y8c ff9 fs1 fc2 sc2 ls0 ws0">：</div><div class="t m0 x135 h3 y8d ff2 fs1 fc2 sc2 ls0 ws0">２<span class="_ _4"></span>０<span class="_ _4"></span>１<span class="_ _4"></span>７</div><div class="t m1 x17 h3 y8e ff2 fs1 fc2 sc2 ls0 ws0">－</div><div class="t m0 xff h3 y8d ff2 fs1 fc2 sc2 ls0 ws0">０<span class="_ _4"></span>２</div><div class="t m1 xdd h3 y8e ff2 fs1 fc2 sc2 ls0 ws0">－</div><div class="t m0 x145 h3 y8d ff2 fs1 fc2 sc2 ls0 ws0">０<span class="_ _4"></span>８</div></div><div class="pi" data-data='{"ctm":[1.830664,0.000000,0.000000,1.830664,0.000000,0.000000]}'></div></div>
</body>
</html>
<div id="pf2" class="pf w0 h0" data-page-no="2"><div class="pc pc2 w0 h0"><img class="bi x0 y0 w1 h1" alt="" src="https://static.pudn.com/prod/directory_preview_static/625fb7df8cbeb85d57682fe6/bg2.jpg"><div class="t m0 x57 hd y8f ff8 fs5 fc2 sc2 ls4 ws0">语音信号是一种非平稳的随机信号</div><div class="t m0 x80 h16 y90 ffd fs5 fc2 sc2 ls0 ws0">，</div><div class="t m0 x13e hd y8f ff8 fs5 fc2 sc2 ls4 ws0">其形成和感知的过程就是一个复杂信号的处理过程</div><div class="t m0 x146 h16 y90 ffd fs5 fc2 sc2 ls0 ws0">。</div><div class="t m0 x49 hd y8f ff8 fs5 fc2 sc2 ls0 ws0">同<span class="_ _e"></span>时</div><div class="t m0 xda h16 y90 ffd fs5 fc2 sc2 ls0 ws0">，</div><div class="t m0 xc1 hd y8f ff8 fs5 fc2 sc2 ls0 ws0">人</div><div class="t m0 x27 hd y91 ff8 fs5 fc2 sc2 ls4 ws0">类大脑是一种多层或者深层处理结构</div><div class="t m0 xb1 h16 y92 ffd fs5 fc2 sc2 ls0 ws0">，</div><div class="t m0 x97 hd y91 ff8 fs5 fc2 sc2 ls6 ws0">对<span class="_ _d"></span>语音<span class="_ _16"></span>信号的<span class="_ _d"></span>处理是一种层次化的处理过程</div><div class="t m0 x147 h16 y92 ffd fs5 fc2 sc2 ls0 ws0">。</div><div class="t m0 xfa hd y91 ff8 fs5 fc2 sc2 ls6 ws0">浅层模型在语音</div><div class="t m0 x27 hd y93 ff8 fs5 fc2 sc2 ls4 ws0">信号的处理过程中会相对受限</div><div class="t m0 xc8 h16 y94 ffd fs5 fc2 sc2 ls0 ws0">，</div><div class="t m0 xc9 hd y93 ff8 fs5 fc2 sc2 ls4 ws0">而深层模型在一<span class="ls6">定程度<span class="_ _d"></span>上模拟人类语音信息的结构化提取过程</span></div><div class="t m0 x148 h16 y94 ffd fs5 fc2 sc2 ls0 ws0">。</div><div class="t m0 x149 hd y93 ff8 fs5 fc2 sc2 ls0 ws0">由<span class="_ _e"></span>此</div><div class="t m0 x27 hd y95 ff8 fs5 fc2 sc2 ls0 ws0">可<span class="_ _f"></span>见</div><div class="t m0 x57 h16 y96 ffd fs5 fc2 sc2 ls0 ws0">，</div><div class="t m0 xa0 hd y95 ff8 fs5 fc2 sc2 ls4 ws0">深层模型比浅层模型更适合于语音信号处理</div><div class="t m0 xf3 h16 y96 ffd fs5 fc2 sc2 ls0 ws0">。</div><div class="t m0 x63 hd y95 ff8 fs5 fc2 sc2 ls6 ws0">深度学<span class="_ _d"></span>习的优势吸引了很多语音信号处理领域的</div><div class="t m0 x27 hd y97 ff8 fs5 fc2 sc2 ls4 ws0">研究人员的关注</div><div class="t m0 xa2 h16 y98 ffd fs5 fc2 sc2 ls0 ws0">，</div><div class="t m0 x25 hd y97 ff8 fs5 fc2 sc2 ls4 ws0">人们开始对其展开了积极的研究</div><div class="t m0 xf3 h16 y98 ffd fs5 fc2 sc2 ls0 ws0">。</div><div class="t m0 x63 hd y97 ff8 fs5 fc2 sc2 ls0 ws0">在<span class="_ _e"></span>后<span class="_ _e"></span>来<span class="_ _17"></span>的<span class="_ _e"></span>几<span class="_ _e"></span>年<span class="_ _17"></span>里</div><div class="t m0 x118 h16 y98 ffd fs5 fc2 sc2 ls0 ws0">，</div><div class="t m0 x14a hd y97 ff8 fs5 fc2 sc2 ls6 ws0">经过研究人员的不懈努力</div><div class="t m0 x14b h16 y98 ffd fs5 fc2 sc2 ls0 ws0">，</div><div class="t m0 x149 hd y97 ff8 fs5 fc2 sc2 ls0 ws0">取<span class="_ _e"></span>得</div><div class="t m0 x27 hd y99 ff8 fs5 fc2 sc2 ls4 ws0">了很多突破性的进展</div><div class="t m0 x12 h16 y9a ffd fs5 fc2 sc2 ls0 ws0">。</div><div class="t m0 x6b h11 y9b ff2 fs5 fc2 sc2 ls0 ws0">２<span class="_ _7"></span>０<span class="_ _8"></span>０<span class="_ _8"></span>９</div><div class="t m0 x134 hd y99 ff8 fs5 fc2 sc2 ls0 ws0">年</div><div class="t m0 x2f h16 y9a ffd fs5 fc2 sc2 ls0 ws0">，</div><div class="t m0 x14c hd y99 ff8 fs5 fc2 sc2 ls4 ws0">深度学习首次被<span class="ls6">应用到语音识别任务</span></div><div class="t m0 x14d h19 y9c ffd fs7 fc2 sc2 ls0 ws0">［</div><div class="t m0 x14e h1a y9d ff2 fs7 fc2 sc2 ls0 ws0">５</div><div class="t m0 xe1 h19 y9c ffd fs7 fc2 sc2 ls0 ws0">］</div><div class="t m0 xe2 h16 y9a ffd fs5 fc2 sc2 ls0 ws0">，</div><div class="t m0 x9d hd y99 ff8 fs5 fc2 sc2 ls6 ws0">相比于传统的高斯混合模型</div><div class="t m4 xeb h11 y9e ff2 fs5 fc2 sc2 ls0 ws0">－</div><div class="t m0 x27 hd y9f ff8 fs5 fc2 sc2 ls0 ws0">隐<span class="_ _f"></span>马<span class="_ _f"></span>尔<span class="_ _f"></span>科<span class="_ _e"></span>夫<span class="_ _f"></span>模<span class="_ _f"></span>型</div><div class="t m0 x70 h16 ya0 ffd fs5 fc2 sc2 ls0 ws0">（</div><div class="t m0 x4e h11 ya1 ff2 fs5 fc2 sc2 ls0 ws0">Ｇ<span class="_ _4"></span>ａ<span class="_ _7"></span>ｕ<span class="_ _9"></span>ｓ<span class="_ _9"></span>ｓ<span class="_ _13"></span>ｉ<span class="_ _13"></span>ａ<span class="_ _7"></span>ｎ</div><div class="t m0 x14f he ya2 ff3 fs5 fc2 sc2 ls0 ws0">　</div><div class="t m0 x88 h11 ya1 ff2 fs5 fc2 sc2 ls0 ws0">ｍ<span class="_ _1"></span>ｉ<span class="_ _b"></span>ｘ<span class="_ _a"></span>ｔ<span class="_ _9"></span>ｕ<span class="_ _9"></span>ｒ<span class="_ _a"></span>ｅ</div><div class="t m0 x95 he ya2 ff3 fs5 fc2 sc2 ls0 ws0">　</div><div class="t m0 x7f h11 ya1 ff2 fs5 fc2 sc2 ls0 ws0">ｍ<span class="_ _6"></span>ｏ<span class="_ _7"></span>ｄ<span class="_ _7"></span>ｅ<span class="_ _b"></span>ｌ</div><div class="t m4 x13e h11 ya3 ff2 fs5 fc2 sc2 ls0 ws0">－</div><div class="t m0 x32 h11 ya1 ff2 fs5 fc2 sc2 ls0 ws0">ｈ<span class="_ _b"></span>ｉ<span class="_ _b"></span>ｄ<span class="_ _8"></span>ｄ<span class="_ _9"></span>ｅ<span class="_ _7"></span>ｎ</div><div class="t m0 x150 he ya2 ff3 fs5 fc2 sc2 ls0 ws0">　</div><div class="t m0 x5d h11 ya1 ff2 fs5 fc2 sc2 ls0 ws0">Ｍ<span class="_ _6"></span>ａ<span class="_ _9"></span>ｒ<span class="_ _9"></span>ｋ<span class="_ _8"></span>ｏ<span class="_ _7"></span>ｖ</div><div class="t m0 x13a he ya2 ff3 fs5 fc2 sc2 ls0 ws0">　</div><div class="t m0 x151 h11 ya1 ff2 fs5 fc2 sc2 ls0 ws0">ｍ<span class="_ _6"></span>ｏ<span class="_ _7"></span>ｄ<span class="_ _7"></span>ｅ<span class="_ _b"></span>ｌ</div><div class="t m0 x13d h16 ya0 ffd fs5 fc2 sc2 ls0 ws0">，</div><div class="t m0 x5b h11 ya1 ff2 fs5 fc2 sc2 ls0 ws0">Ｇ<span class="_ _0"></span>ＭＭ</div><div class="t m4 x152 h11 ya2 ff2 fs5 fc2 sc2 ls0 ws0">－</div><div class="t m0 x105 h11 ya1 ff2 fs5 fc2 sc2 ls0 ws0">Ｈ<span class="_ _d"></span>ＭＭ</div><div class="t m0 x153 h16 ya0 ffd fs5 fc2 sc2 ls0 ws0">）</div><div class="t m0 xf9 hd y9f ff8 fs5 fc2 sc2 ls5 ws0">语音识别系统获得了超</div><div class="t m0 x27 hd ya4 ff8 fs5 fc2 sc2 ls0 ws0">过</div><div class="t m0 x56 h11 ya5 ff2 fs5 fc2 sc2 ls0 ws0">２<span class="_ _7"></span>０<span class="_ _c"></span>％</div><div class="t m0 xd hd ya4 ff8 fs5 fc2 sc2 ls4 ws0">的相对性能提升</div><div class="t m0 xae h16 ya6 ffd fs5 fc2 sc2 ls0 ws0">。</div><div class="t m0 x92 hd ya4 ff8 fs5 fc2 sc2 ls0 ws0">此<span class="_ _f"></span>后</div><div class="t m0 x2f h16 ya6 ffd fs5 fc2 sc2 ls0 ws0">，</div><div class="t m0 x14c hd ya4 ff8 fs5 fc2 sc2 ls4 ws0">基于深度神经网络</div><div class="t m0 xde h16 ya6 ffd fs5 fc2 sc2 ls0 ws0">（</div><div class="t m0 x154 h11 ya5 ff2 fs5 fc2 sc2 ls0 ws0">Ｄ<span class="_ _1"></span>ｅ<span class="_ _9"></span>ｅ</div><div class="t m0 x155 h11 ya7 ff2 fs5 fc2 sc2 ls0 ws0">ｐ</div><div class="t m0 x139 he ya8 ff3 fs5 fc2 sc2 ls0 ws0">　</div><div class="t m0 x156 h11 ya5 ff2 fs5 fc2 sc2 ls0 ws0">ｎ<span class="_ _7"></span>ｅ<span class="_ _7"></span>ｕ<span class="_ _9"></span>ｒ<span class="_ _a"></span>ａ<span class="_ _b"></span>ｌ</div><div class="t m0 x5f he ya9 ff3 fs5 fc2 sc2 ls0 ws0">　</div><div class="t m0 xdf h11 ya5 ff2 fs5 fc2 sc2 ls0 ws0">ｎ<span class="_ _7"></span>ｅ<span class="_ _b"></span>ｔ<span class="_ _1"></span>ｗ<span class="_ _2"></span>ｏ<span class="_ _a"></span>ｒ<span class="_ _7"></span>ｋ<span class="_ _7"></span>ｓ</div><div class="t m0 x157 h16 ya6 ffd fs5 fc2 sc2 ls0 ws0">，</div><div class="t m0 xb7 h11 ya5 ff2 fs5 fc2 sc2 ls0 ws0">Ｄ<span class="_ _3"></span>Ｎ<span class="_ _c"></span>Ｎ</div><div class="t m0 x108 h16 ya6 ffd fs5 fc2 sc2 ls0 ws0">）</div><div class="t m0 xbb hd ya4 ff8 fs5 fc2 sc2 ls5 ws0">的声学模型逐渐替代</div><div class="t m0 x27 hd yaa ff8 fs5 fc2 sc2 ls0 ws0">了</div><div class="t m0 xc2 h11 yab ff2 fs5 fc2 sc2 ls0 ws0">Ｇ<span class="_ _0"></span>ＭＭ</div><div class="t m0 xaa hd yaa ff8 fs5 fc2 sc2 ls4 ws0">成为语音识别声学建模的主流模型</div><div class="t m0 x158 h16 yac ffd fs5 fc2 sc2 ls0 ws0">，</div><div class="t m0 x82 hd yaa ff8 fs5 fc2 sc2 ls4 ws0">并极大地促进了语音识别技术的发展</div><div class="t m0 x61 h16 yac ffd fs5 fc2 sc2 ls0 ws0">，</div><div class="t m0 x3d hd yaa ff8 fs5 fc2 sc2 ls0 ws0">突<span class="_ _f"></span>破<span class="_ _f"></span>了<span class="_ _e"></span>某<span class="_ _f"></span>些<span class="_ _e"></span>实<span class="_ _e"></span>际<span class="_ _e"></span>应</div><div class="t m0 x27 hd yad ff8 fs5 fc2 sc2 ls4 ws0">用场景下对语音识别性能要求的瓶颈</div><div class="t m0 xb1 h16 yae ffd fs5 fc2 sc2 ls0 ws0">，</div><div class="t m0 x97 hd yad ff8 fs5 fc2 sc2 ls0 ws0">使<span class="_ _f"></span>语<span class="_ _f"></span>音<span class="_ _e"></span>识<span class="_ _f"></span>别<span class="_ _e"></span>技<span class="_ _e"></span>术<span class="_ _e"></span>走<span class="_ _e"></span>向<span class="_ _17"></span>真<span class="_ _e"></span>正<span class="_ _e"></span>实<span class="_ _17"></span>用<span class="_ _e"></span>化</div><div class="t m0 x105 h16 yae ffd fs5 fc2 sc2 ls0 ws0">。</div><div class="t m0 xd6 hd yad ff8 fs5 fc2 sc2 ls6 ws0">本文首先对深度学习做一个</div><div class="t m0 x27 hd yaf ff8 fs5 fc2 sc2 ls0 ws0">简<span class="_ _f"></span>要<span class="_ _f"></span>概<span class="_ _f"></span>述</div><div class="t m0 xaa h16 yb0 ffd fs5 fc2 sc2 ls0 ws0">，</div><div class="t m0 x86 hd yaf ff8 fs5 fc2 sc2 ls4 ws0">然后重点就基于深度学习的语音识别技术现状进行较为详细的讨论</div><div class="t m0 xba h16 yb0 ffd fs5 fc2 sc2 ls0 ws0">，</div><div class="t m0 xd7 hd yaf ff8 fs5 fc2 sc2 ls4 ws0">最后就基于深度学<span class="_ _10"></span>习的</div><div class="t m0 x27 hd yb1 ff8 fs5 fc2 sc2 ls4 ws0">语音识别未来可能的研究方向进行展望</div><div class="t m0 x80 h16 yb2 ffd fs5 fc2 sc2 ls0 ws0">。</div><div class="t m0 x159 h12 yb3 ffc fs6 fc2 sc2 ls0 ws0">１</div><div class="t m0 x15a h13 yb4 ff3 fs6 fc2 sc2 ls0 ws0">　</div><div class="t m0 x4 h18 yb5 ff6 fs6 fc2 sc2 ls0 ws0">深<span class="_ _f"></span>度<span class="_ _f"></span>学<span class="_ _e"></span>习<span class="_ _f"></span>简<span class="_ _f"></span>介</div><div class="t m0 x57 hd yb6 ff8 fs5 fc2 sc2 ls4 ws0">深度学习的概念来源于人<span class="_ _10"></span>工<span class="_ _10"></span>神<span class="_ _10"></span>经网<span class="_ _10"></span>络</div><div class="t m0 x6d h16 yb7 ffd fs5 fc2 sc2 ls0 ws0">（</div><div class="t m0 x113 h11 yb8 ff2 fs5 fc2 sc2 ls0 ws0">Ａ<span class="_ _1"></span>ｒ<span class="_ _b"></span>ｔ<span class="_ _13"></span>ｉ<span class="_ _11"></span>ｆ<span class="_ _13"></span>ｉ<span class="_ _13"></span>ｃ<span class="_ _13"></span>ｉ<span class="_ _13"></span>ａ<span class="_ _b"></span>ｌ</div><div class="t m0 x15b he yb9 ff3 fs5 fc2 sc2 ls0 ws0">　</div><div class="t m0 x15c h11 yb8 ff2 fs5 fc2 sc2 ls0 ws0">ｎ<span class="_ _7"></span>ｅ<span class="_ _7"></span>ｕ<span class="_ _9"></span>ｒ<span class="_ _a"></span>ａ<span class="_ _b"></span>ｌ</div><div class="t m0 x15d he yb9 ff3 fs5 fc2 sc2 ls0 ws0">　</div><div class="t m0 x137 h11 yb8 ff2 fs5 fc2 sc2 ls0 ws0">ｎ<span class="_ _7"></span>ｅ<span class="_ _b"></span>ｔ<span class="_ _1"></span>ｗ<span class="_ _2"></span>ｏ<span class="_ _a"></span>ｒ<span class="_ _7"></span>ｋ</div><div class="t m0 xd2 h16 yb7 ffd fs5 fc2 sc2 ls0 ws0">，</div><div class="t m0 x14d h11 yb8 ff2 fs5 fc2 sc2 ls0 ws0">Ａ<span class="_ _5"></span>Ｎ<span class="_ _5"></span>Ｎ</div><div class="t m0 x131 h16 yb7 ffd fs5 fc2 sc2 ls0 ws0">）<span class="_ _8"></span>。</div><div class="t m0 x15e hd yb6 ff8 fs5 fc2 sc2 ls6 ws0">人工神经网络是机器学</div><div class="t m0 x27 hd yba ff8 fs5 fc2 sc2 ls4 ws0">习与人工智能领域的一种模型</div><div class="t m0 xc8 h19 ybb ffd fs7 fc2 sc2 ls0 ws0">［</div><div class="t m0 x74 h1a ybc ff2 fs7 fc2 sc2 ls0 ws0">６</div><div class="t m0 xc9 h19 ybb ffd fs7 fc2 sc2 ls0 ws0">］</div><div class="t m0 x123 h16 ybd ffd fs5 fc2 sc2 ls0 ws0">。</div><div class="t m0 x8a hd yba ff8 fs5 fc2 sc2 ls4 ws0">它从信息处理<span class="ls6">角度对人类大脑的神经元网络进行抽象</span></div><div class="t m0 x37 h16 ybd ffd fs5 fc2 sc2 ls0 ws0">，</div><div class="t m0 x39 hd yba ff8 fs5 fc2 sc2 ls0 ws0">从<span class="_ _e"></span>而<span class="_ _e"></span>达<span class="_ _17"></span>到<span class="_ _e"></span>模</div><div class="t m0 x27 hd ybe ff8 fs5 fc2 sc2 ls4 ws0">拟人脑认知和学习能力的目的</div><div class="t m0 xc8 h16 ybf ffd fs5 fc2 sc2 ls0 ws0">。</div><div class="t m0 x123 hd ybe ff8 fs5 fc2 sc2 ls0 ws0">第</div><div class="t m0 x8a h11 yc0 ff2 fs5 fc2 sc2 ls0 ws0">１</div><div class="t m0 x76 hd ybe ff8 fs5 fc2 sc2 ls4 ws0">个人工神经元叫做阈值逻辑<span class="_ _10"></span>单元</div><div class="t m0 xf7 h16 ybf ffd fs5 fc2 sc2 ls0 ws0">（</div><div class="t m0 x152 h11 yc0 ff2 fs5 fc2 sc2 ls0 ws0">Ｔ<span class="_ _2"></span>ｈ<span class="_ _9"></span>ｒ<span class="_ _a"></span>ｅ<span class="_ _9"></span>ｓ<span class="_ _9"></span>ｈ<span class="_ _8"></span>ｏ<span class="_ _b"></span>ｌ<span class="_ _a"></span>ｄ</div><div class="t m0 x15f he yc1 ff3 fs5 fc2 sc2 ls0 ws0">　</div><div class="t m0 xbb h11 yc0 ff2 fs5 fc2 sc2 ls0 ws0">ｌ<span class="_ _b"></span>ｏ</div><div class="t m0 x160 h11 yc2 ff2 fs5 fc2 sc2 ls0 ws0">ｇ</div><div class="t m0 x3d h11 yc0 ff2 fs5 fc2 sc2 ls0 ws0">ｉ<span class="_ _13"></span>ｃ</div><div class="t m0 xfa he yc1 ff3 fs5 fc2 sc2 ls0 ws0">　</div><div class="t m0 x161 h11 yc0 ff2 fs5 fc2 sc2 ls0 ws0">ｕ<span class="_ _8"></span>ｎ<span class="_ _b"></span>ｉ<span class="_ _13"></span>ｔ</div><div class="t m0 x47 h16 ybf ffd fs5 fc2 sc2 ls0 ws0">，</div><div class="t m0 x162 h11 yc0 ff2 fs5 fc2 sc2 ls0 ws0">Ｔ<span class="_ _6"></span>Ｌ<span class="_ _5"></span>Ｕ</div><div class="t m0 x149 h16 ybf ffd fs5 fc2 sc2 ls0 ws0">）<span class="_ _8"></span>，</div><div class="t m0 xc1 hd ybe ff8 fs5 fc2 sc2 ls0 ws0">是</div><div class="t m0 x27 hd yc3 ff8 fs5 fc2 sc2 ls0 ws0">由</div><div class="t m0 x84 h11 yc4 ff2 fs5 fc2 sc2 ls0 ws0">Ｍ<span class="_ _6"></span>ｃ<span class="_ _1"></span>Ｃ<span class="_ _1"></span>ｕ<span class="_ _a"></span>ｌ<span class="_ _12"></span>ｌ<span class="_ _b"></span>ｏ<span class="_ _9"></span>ｃ<span class="_ _7"></span>ｈ</div><div class="t m0 x163 hd yc3 ff8 fs5 fc2 sc2 ls0 ws0">和</div><div class="t m0 xa2 h11 yc4 ff2 fs5 fc2 sc2 ls0 ws0">Ｐ<span class="_ _a"></span>ｉ<span class="_ _13"></span>ｔ<span class="_ _13"></span>ｔ<span class="_ _b"></span>ｓ</div><div class="t m0 xee hd yc3 ff8 fs5 fc2 sc2 ls0 ws0">在</div><div class="t m0 x121 h11 yc4 ff2 fs5 fc2 sc2 ls0 ws0">１<span class="_ _7"></span>９<span class="_ _8"></span>４<span class="_ _8"></span>３</div><div class="t m0 x7d hd yc3 ff8 fs5 fc2 sc2 ls0 ws0">年<span class="_ _f"></span>提<span class="_ _f"></span>出<span class="_ _e"></span>的</div><div class="t m0 xb2 h16 yc5 ffd fs5 fc2 sc2 ls0 ws0">，</div><div class="t m0 x81 hd yc3 ff8 fs5 fc2 sc2 ls0 ws0">由<span class="_ _e"></span>此</div><div class="t m0 x164 h16 yc5 ffd fs5 fc2 sc2 ls0 ws0">，</div><div class="t m0 x78 hd yc3 ff8 fs5 fc2 sc2 ls8 ws0">开创了人工神经网络研究时代</div><div class="t m0 x165 h16 yc5 ffd fs5 fc2 sc2 ls0 ws0">。</div><div class="t m0 x45 h11 yc4 ff2 fs5 fc2 sc2 ls0 ws0">１<span class="_ _7"></span>９<span class="_ _8"></span>５<span class="_ _8"></span>８</div><div class="t m0 x143 hd yc3 ff8 fs5 fc2 sc2 ls0 ws0">年</div><div class="t m0 x3f h16 yc5 ffd fs5 fc2 sc2 ls0 ws0">，</div><div class="t m0 x166 h11 yc4 ff2 fs5 fc2 sc2 ls0 ws0">Ｒ<span class="_ _4"></span>ｏ<span class="_ _9"></span>ｓ<span class="_ _a"></span>ｅ<span class="_ _7"></span>ｎ<span class="_ _1"></span>ｂ<span class="_ _b"></span>ｌ<span class="_ _b"></span>ａ<span class="_ _b"></span>ｔ<span class="_ _b"></span>ｔ</div><div class="t m0 xc1 hd yc3 ff8 fs5 fc2 sc2 ls0 ws0">提</div><div class="t m0 x27 hd yc6 ff8 fs5 fc2 sc2 ls0 ws0">出<span class="_ _f"></span>了<span class="_ _f"></span>感<span class="_ _f"></span>知<span class="_ _e"></span>器<span class="_ _f"></span>模<span class="_ _f"></span>型</div><div class="t m0 x70 h19 yc7 ffd fs7 fc2 sc2 ls0 ws0">［</div><div class="t m0 x87 h1a yc8 ff2 fs7 fc2 sc2 ls0 ws0">７</div><div class="t m0 x25 h19 yc7 ffd fs7 fc2 sc2 ls0 ws0">］</div><div class="t m0 xc4 h16 yc9 ffd fs5 fc2 sc2 ls0 ws0">，</div><div class="t m0 x4f hd yc6 ff8 fs5 fc2 sc2 ls4 ws0">该模型是使用线性阈值函数的一个非常重要的人工神经网络</div><div class="t m0 x167 h16 yc9 ffd fs5 fc2 sc2 ls0 ws0">，</div><div class="t m0 x168 hd yc6 ff8 fs5 fc2 sc2 ls5 ws0">在人<span class="_ _d"></span>工神经网络的发</div><div class="t m0 x27 hd yca ff8 fs5 fc2 sc2 ls4 ws0">展过程中具有开拓性意义</div><div class="t m0 x65 h16 ycb ffd fs5 fc2 sc2 ls0 ws0">。</div><div class="t m0 x66 hd yca ff8 fs5 fc2 sc2 ls0 ws0">另<span class="_ _f"></span>外</div><div class="t m0 x123 h16 ycb ffd fs5 fc2 sc2 ls0 ws0">，</div><div class="t m0 x17 h11 ycc ff2 fs5 fc2 sc2 ls0 ws0">Ｒ<span class="_ _4"></span>ｏ<span class="_ _9"></span>ｓ<span class="_ _a"></span>ｅ<span class="_ _7"></span>ｎ<span class="_ _1"></span>ｂ<span class="_ _b"></span>ｌ<span class="_ _b"></span>ａ<span class="_ _b"></span>ｔ<span class="_ _b"></span>ｔ</div><div class="t m0 x169 hd yca ff8 fs5 fc2 sc2 ls4 ws0">还通过模拟人类学习的过程</div><div class="t m0 xe3 h16 ycb ffd fs5 fc2 sc2 ls0 ws0">，</div><div class="t m0 x16a hd yca ff8 fs5 fc2 sc2 ls0 ws0">在</div><div class="t m0 x16b h11 ycc ff2 fs5 fc2 sc2 ls0 ws0">Ｈ<span class="_ _6"></span>ｅ<span class="_ _9"></span>ｂ<span class="_ _1"></span>ｂ</div><div class="t m0 x61 hd yca ff8 fs5 fc2 sc2 ls0 ws0">学<span class="_ _f"></span>习<span class="_ _f"></span>法<span class="_ _e"></span>则<span class="_ _f"></span>的<span class="_ _e"></span>基<span class="_ _e"></span>础<span class="_ _e"></span>上</div><div class="t m0 xeb h16 ycb ffd fs5 fc2 sc2 ls0 ws0">，</div><div class="t m0 x27 hd ycd ff8 fs5 fc2 sc2 ls0 ws0">提<span class="_ _f"></span>出<span class="_ _f"></span>了<span class="_ _f"></span>一<span class="_ _e"></span>种<span class="_ _f"></span>迭<span class="_ _f"></span>代</div><div class="t m0 xa2 h16 yce ffd fs5 fc2 sc2 ls0 ws0">、</div><div class="t m0 x2c hd ycd ff8 fs5 fc2 sc2 ls4 ws0">试错的学习算法</div><div class="t m0 xff h16 yce ffd fs5 fc2 sc2 ls0 ws0">—<span class="_ _7"></span>—<span class="_ _8"></span>—</div><div class="t m0 x97 hd ycd ff8 fs5 fc2 sc2 ls0 ws0">感<span class="_ _f"></span>知<span class="_ _f"></span>器<span class="_ _e"></span>学<span class="_ _e"></span>习<span class="_ _17"></span>算<span class="_ _e"></span>法</div><div class="t m0 x155 h16 yce ffd fs5 fc2 sc2 ls0 ws0">。</div><div class="t m0 x13b hd ycd ff8 fs5 fc2 sc2 ls0 ws0">感<span class="_ _e"></span>知<span class="_ _e"></span>器<span class="_ _17"></span>是<span class="_ _e"></span>第</div><div class="t m0 x118 h11 ycf ff2 fs5 fc2 sc2 ls0 ws0">１</div><div class="t m0 x16c hd ycd ff8 fs5 fc2 sc2 ls6 ws0">个可学习的神经网络模型</div><div class="t m0 x149 h16 yce ffd fs5 fc2 sc2 ls0 ws0">。</div><div class="t m0 xc1 hd ycd ff8 fs5 fc2 sc2 ls0 ws0">这</div><div class="t m0 x27 hd yd0 ff8 fs5 fc2 sc2 ls4 ws0">时感知器一般是指单层非线性变换的网络结构</div><div class="t m0 x78 h16 yd1 ffd fs5 fc2 sc2 ls0 ws0">，</div><div class="t m0 x142 hd yd0 ff8 fs5 fc2 sc2 ls4 ws0">它仅对线性可分问题具有分类能力</div><div class="t m0 x61 h16 yd1 ffd fs5 fc2 sc2 ls0 ws0">，</div><div class="t m0 x3d hd yd0 ff8 fs5 fc2 sc2 ls0 ws0">对<span class="_ _f"></span>于<span class="_ _f"></span>线<span class="_ _e"></span>性<span class="_ _f"></span>不<span class="_ _e"></span>可<span class="_ _e"></span>分<span class="_ _e"></span>问</div><div class="t m0 x27 hd yd2 ff8 fs5 fc2 sc2 ls4 ws0">题只能做近似分类</div><div class="t m0 x5c h16 yd3 ffd fs5 fc2 sc2 ls0 ws0">。</div><div class="t m0 x12 hd yd2 ff8 fs5 fc2 sc2 ls0 ws0">多<span class="_ _f"></span>层<span class="_ _f"></span>感<span class="_ _e"></span>知<span class="_ _f"></span>器</div><div class="t m0 x123 h16 yd3 ffd fs5 fc2 sc2 ls0 ws0">（</div><div class="t m0 xff h11 yd4 ff2 fs5 fc2 sc2 ls0 ws0">Ｍ<span class="_ _3"></span>ｕ<span class="_ _a"></span>ｌ<span class="_ _13"></span>ｔ<span class="_ _12"></span>ｉ<span class="_ _11"></span>ｌ<span class="_ _b"></span>ａ</div><div class="t m0 x81 h11 yd5 ff2 fs5 fc2 sc2 ls0 ws0">ｙ</div><div class="t m0 x32 h11 yd4 ff2 fs5 fc2 sc2 ls0 ws0">ｅ<span class="_ _a"></span>ｒ</div><div class="t m0 x113 he yd6 ff3 fs5 fc2 sc2 ls0 ws0">　</div><div class="t m0 x158 h11 yd5 ff2 fs5 fc2 sc2 ls0 ws0">ｐ</div><div class="t m0 x164 h11 yd4 ff2 fs5 fc2 sc2 ls0 ws0">ｅ<span class="_ _a"></span>ｒ<span class="_ _a"></span>ｃ<span class="_ _a"></span>ｅ</div><div class="t m0 x16d h11 yd5 ff2 fs5 fc2 sc2 ls0 ws0">ｐ</div><div class="t m0 x102 h11 yd4 ff2 fs5 fc2 sc2 ls0 ws0">ｔ<span class="_ _b"></span>ｒ<span class="_ _a"></span>ｏ<span class="_ _8"></span>ｎ</div><div class="t m0 x155 h16 yd3 ffd fs5 fc2 sc2 ls0 ws0">，</div><div class="t m0 x156 h11 yd4 ff2 fs5 fc2 sc2 ls0 ws0">Ｍ<span class="_ _c"></span>Ｌ<span class="_ _2"></span>Ｐ</div><div class="t m0 x138 h16 yd3 ffd fs5 fc2 sc2 ls0 ws0">）</div><div class="t m0 x13d hd yd2 ff8 fs5 fc2 sc2 ls6 ws0">解决了单层感知器的局限性问题</div><div class="t m0 x14b h16 yd3 ffd fs5 fc2 sc2 ls0 ws0">，</div><div class="t m0 x149 hd yd2 ff8 fs5 fc2 sc2 ls0 ws0">利<span class="_ _e"></span>用</div><div class="t m0 x27 hd yd7 ff8 fs5 fc2 sc2 ls4 ws0">多个单层感知器堆叠而成</div><div class="t m0 x65 h16 yd8 ffd fs5 fc2 sc2 ls0 ws0">，</div><div class="t m0 x88 hd yd7 ff8 fs5 fc2 sc2 ls4 ws0">并且采用的激活函数不是线性阈值函数而是连续非线性函数</div><div class="t m0 x3e h16 yd8 ffd fs5 fc2 sc2 ls0 ws0">。</div><div class="t m0 x42 hd yd7 ff8 fs5 fc2 sc2 ls0 ws0">由<span class="_ _f"></span>于</div><div class="t m0 x4a h11 yd9 ff2 fs5 fc2 sc2 ls0 ws0">Ｍ<span class="_ _c"></span>Ｌ<span class="_ _2"></span>Ｐ</div><div class="t m0 xc1 hd yd7 ff8 fs5 fc2 sc2 ls0 ws0">是</div><div class="t m0 x27 hd yda ff8 fs5 fc2 sc2 ls4 ws0">一种多层的非线性变换模型</div><div class="t m0 x66 h16 ydb ffd fs5 fc2 sc2 ls0 ws0">，</div><div class="t m0 x94 hd yda ff8 fs5 fc2 sc2 ls6 ws0">其<span class="_ _d"></span>具有<span class="_ _16"></span>强大的表达和建模能力</div><div class="t m0 x16e h16 ydb ffd fs5 fc2 sc2 ls0 ws0">。</div><div class="t m0 x16f hd yda ff8 fs5 fc2 sc2 ls0 ws0">同<span class="_ _e"></span>时</div><div class="t m0 x170 h16 ydb ffd fs5 fc2 sc2 ls0 ws0">，</div><div class="t m0 xb6 h11 ydc ff2 fs5 fc2 sc2 ls0 ws0">Ｍ<span class="_ _c"></span>Ｌ<span class="_ _2"></span>Ｐ</div><div class="t m0 x16a hd yda ff8 fs5 fc2 sc2 ls6 ws0">可以通过误差后向传播算法</div><div class="t m0 x27 h16 ydd ffd fs5 fc2 sc2 ls0 ws0">（</div><div class="t m0 x8d h11 yde ff2 fs5 fc2 sc2 ls0 ws0">Ｂ<span class="_ _8"></span>ａ<span class="_ _9"></span>ｃ<span class="_ _7"></span>ｋ</div><div class="t m0 xb he y62 ff3 fs5 fc2 sc2 ls0 ws0">　</div><div class="t m0 x6f h11 ydf ff2 fs5 fc2 sc2 ls0 ws0">ｐ</div><div class="t m0 x29 h11 ye0 ff2 fs5 fc2 sc2 ls0 ws0">ｒ<span class="_ _9"></span>ｏ</div><div class="t m0 xaa h11 ydf ff2 fs5 fc2 sc2 ls0 ws0">ｐ</div><div class="t m0 x86 h11 ye0 ff2 fs5 fc2 sc2 ls0 ws0">ａ</div><div class="t m0 x171 h11 ydf ff2 fs5 fc2 sc2 ls0 ws0">ｇ</div><div class="t m0 xf h11 ye0 ff2 fs5 fc2 sc2 ls0 ws0">ａ<span class="_ _b"></span>ｔ<span class="_ _13"></span>ｉ<span class="_ _b"></span>ｏ<span class="_ _7"></span>ｎ</div><div class="t m0 x26 h16 ydd ffd fs5 fc2 sc2 ls0 ws0">，</div><div class="t m0 x68 h11 yde ff2 fs5 fc2 sc2 ls0 ws0">Ｂ<span class="_ _4"></span>Ｐ</div><div class="t m0 x172 h16 ydd ffd fs5 fc2 sc2 ls0 ws0">）</div><div class="t m0 xae h19 ye1 ffd fs7 fc2 sc2 ls0 ws0">［</div><div class="t m0 x121 h1a ye2 ff2 fs7 fc2 sc2 ls0 ws0">８</div><div class="t m0 x15 h19 ye1 ffd fs7 fc2 sc2 ls0 ws0">］</div><div class="t m0 x14f hd ye3 ff8 fs5 fc2 sc2 ls0 ws0">进<span class="_ _f"></span>行<span class="_ _f"></span>训<span class="_ _e"></span>练</div><div class="t m0 x173 h16 ydd ffd fs5 fc2 sc2 ls0 ws0">。</div><div class="t m0 x112 hd ye3 ff8 fs5 fc2 sc2 ls0 ws0">但<span class="_ _f"></span>由<span class="_ _f"></span>于</div><div class="t m0 xb3 h11 yde ff2 fs5 fc2 sc2 ls0 ws0">Ｍ<span class="_ _c"></span>Ｌ<span class="_ _2"></span>Ｐ</div><div class="t m0 xde hd ye3 ff8 fs5 fc2 sc2 ls8 ws0">的各层激活函数均为非线性函数</div><div class="t m0 x174 h16 ydd ffd fs5 fc2 sc2 ls0 ws0">，</div><div class="t m0 xfa hd ye3 ff8 fs5 fc2 sc2 ls8 ws0">模型训练中的损</div><div class="t m0 x27 hd ye4 ff8 fs5 fc2 sc2 ls4 ws0">失函数是模型参数的非凸复杂函数</div><div class="t m0 x175 h16 ye5 ffd fs5 fc2 sc2 ls0 ws0">。</div><div class="t m0 x76 hd ye4 ff8 fs5 fc2 sc2 ls0 ws0">并<span class="_ _f"></span>且</div><div class="t m0 x13f h16 ye5 ffd fs5 fc2 sc2 ls0 ws0">，</div><div class="t m0 x176 hd ye4 ff8 fs5 fc2 sc2 ls0 ws0">随<span class="_ _f"></span>着<span class="_ _f"></span>层<span class="_ _e"></span>数<span class="_ _e"></span>的<span class="_ _e"></span>增<span class="_ _e"></span>多</div><div class="t m0 x137 h16 ye5 ffd fs5 fc2 sc2 ls0 ws0">，</div><div class="t m0 x116 hd ye4 ff8 fs5 fc2 sc2 ls6 ws0">非凸目标函数越来越复杂</div><div class="t m0 xe6 h16 ye5 ffd fs5 fc2 sc2 ls0 ws0">，</div><div class="t m0 xe7 hd ye4 ff8 fs5 fc2 sc2 ls0 ws0">局<span class="_ _e"></span>部<span class="_ _e"></span>最<span class="_ _17"></span>小<span class="_ _e"></span>值<span class="_ _17"></span>点</div><div class="t m0 x27 hd ye6 ff8 fs5 fc2 sc2 ls0 ws0">成<span class="_ _f"></span>倍<span class="_ _f"></span>增<span class="_ _e"></span>长</div><div class="t m0 xa1 h16 ye7 ffd fs5 fc2 sc2 ls0 ws0">，</div><div class="t m0 x86 hd ye6 ff8 fs5 fc2 sc2 ls0 ws0">很<span class="_ _f"></span>难<span class="_ _f"></span>进<span class="_ _e"></span>行<span class="_ _f"></span>优<span class="_ _f"></span>化</div><div class="t m0 x14 h16 ye7 ffd fs5 fc2 sc2 ls0 ws0">，</div><div class="t m0 x65 hd ye6 ff8 fs5 fc2 sc2 ls0 ws0">使<span class="_ _f"></span>用</div><div class="t m0 x74 h11 ye8 ff2 fs5 fc2 sc2 ls0 ws0">Ｂ<span class="_ _4"></span>Ｐ</div><div class="t m0 x31 hd ye6 ff8 fs5 fc2 sc2 ls6 ws0">进行算法进行网络训练时很难<span class="ls8">获得全局最优解</span></div><div class="t m0 x46 h16 ye7 ffd fs5 fc2 sc2 ls0 ws0">。</div><div class="t m0 x174 hd ye6 ff8 fs5 fc2 sc2 ls0 ws0">因<span class="_ _e"></span>此</div><div class="t m0 x37 h16 ye7 ffd fs5 fc2 sc2 ls0 ws0">，</div><div class="t m0 x39 hd ye6 ff8 fs5 fc2 sc2 ls0 ws0">目<span class="_ _e"></span>标<span class="_ _17"> </span>函<span class="_ _17"></span>数<span class="_ _e"></span>难</div><div class="t m0 x27 hd ye9 ff8 fs5 fc2 sc2 ls4 ws0">以优化的问题导致了</div><div class="t m0 x13 h11 yea ff2 fs5 fc2 sc2 ls0 ws0">Ｍ<span class="_ _c"></span>Ｌ<span class="_ _2"></span>Ｐ</div><div class="t m0 xdb hd yeb ff8 fs5 fc2 sc2 ls4 ws0">难以展现其强大的表达和建模能力</div><div class="t m0 x138 h16 yec ffd fs5 fc2 sc2 ls0 ws0">。</div><div class="t m0 x57 hd yed ff8 fs5 fc2 sc2 ls0 ws0">深<span class="_ _f"></span>度<span class="_ _f"></span>置<span class="_ _e"></span>信<span class="_ _f"></span>网<span class="_ _f"></span>络</div><div class="t m0 x5c h16 yee ffd fs5 fc2 sc2 ls0 ws0">（</div><div class="t m0 x68 h11 yef ff2 fs5 fc2 sc2 ls0 ws0">Ｄ<span class="_ _1"></span>ｅ<span class="_ _9"></span>ｅ</div><div class="t m0 x72 h11 yf0 ff2 fs5 fc2 sc2 ls0 ws0">ｐ</div><div class="t m0 x121 he yf1 ff3 fs5 fc2 sc2 ls0 ws0">　</div><div class="t m0 x177 h11 yef ff2 fs5 fc2 sc2 ls0 ws0">ｂ<span class="_ _7"></span>ｅ<span class="_ _b"></span>ｌ<span class="_ _15"></span>ｉ<span class="_ _b"></span>ｅ<span class="_ _b"></span>ｆ</div><div class="t m0 x2e he yf2 ff3 fs5 fc2 sc2 ls0 ws0">　</div><div class="t m0 x7e h11 yef ff2 fs5 fc2 sc2 ls0 ws0">ｎ<span class="_ _7"></span>ｅ<span class="_ _b"></span>ｔ<span class="_ _1"></span>ｗ<span class="_ _2"></span>ｏ<span class="_ _a"></span>ｒ<span class="_ _7"></span>ｋ<span class="_ _7"></span>ｓ</div><div class="t m0 x81 h16 yee ffd fs5 fc2 sc2 ls0 ws0">，</div><div class="t m0 x13f h11 yef ff2 fs5 fc2 sc2 ls0 ws0">Ｄ<span class="_ _6"></span>Ｂ<span class="_ _6"></span>Ｎ</div><div class="t m0 xf1 h16 yee ffd fs5 fc2 sc2 ls0 ws0">）</div><div class="t m0 xf2 h19 yf3 ffd fs7 fc2 sc2 ls0 ws0">［</div><div class="t m0 x178 h1a yf4 ff2 fs7 fc2 sc2 ls0 ws0">１</div><div class="t m0 x16d h19 yf3 ffd fs7 fc2 sc2 ls0 ws0">］</div><div class="t m0 x102 hd yed ff8 fs5 fc2 sc2 ls0 ws0">是</div><div class="t m0 x179 h11 yef ff2 fs5 fc2 sc2 ls0 ws0">Ｈ<span class="_ _1"></span>ｉ<span class="_ _b"></span>ｎ<span class="_ _9"></span>ｔ<span class="_ _a"></span>ｏ<span class="_ _7"></span>ｎ</div><div class="t m0 x17a hd yed ff8 fs5 fc2 sc2 ls0 ws0">等<span class="_ _e"></span>学<span class="_ _e"></span>者<span class="_ _17"></span>在</div><div class="t m0 x14a h11 yef ff2 fs5 fc2 sc2 ls0 ws0">２<span class="_ _7"></span>０<span class="_ _8"></span>０<span class="_ _8"></span>６</div><div class="t m0 x106 hd yed ff8 fs5 fc2 sc2 ls6 ws0">年提出的一种无监督的概</div><div class="t m0 x27 hd yf5 ff8 fs5 fc2 sc2 ls0 ws0">率<span class="_ _f"></span>生<span class="_ _f"></span>成<span class="_ _e"></span>模<span class="_ _f"></span>型</div><div class="t m0 x62 h16 yf6 ffd fs5 fc2 sc2 ls0 ws0">，</div><div class="t m0 x17b hd yf5 ff8 fs5 fc2 sc2 ls0 ws0">用</div><div class="t m0 x11 h11 yf7 ff2 fs5 fc2 sc2 ls0 ws0">Ｄ<span class="_ _6"></span>Ｂ<span class="_ _6"></span>Ｎ</div><div class="t m0 x12 hd yf5 ff8 fs5 fc2 sc2 ls0 ws0">来<span class="_ _f"></span>初<span class="_ _f"></span>始<span class="_ _e"></span>化</div><div class="t m0 x7e h11 yf7 ff2 fs5 fc2 sc2 ls0 ws0">Ｍ<span class="_ _c"></span>Ｌ<span class="_ _2"></span>Ｐ</div><div class="t m0 x145 hd yf5 ff8 fs5 fc2 sc2 ls4 ws0">各层的网络参<span class="ls5">数能够解决其目标函数难以优化的问题</span></div><div class="t m0 x39 h16 yf6 ffd fs5 fc2 sc2 ls0 ws0">。</div><div class="t m0 x11d hd yf5 ff8 fs5 fc2 sc2 ls0 ws0">一<span class="_ _e"></span>般<span class="_ _e"></span>称<span class="_ _17"></span>使</div><div class="t m0 x27 hd yf8 ff8 fs5 fc2 sc2 ls0 ws0">用</div><div class="t m0 xc2 h11 yf9 ff2 fs5 fc2 sc2 ls0 ws0">Ｄ<span class="_ _6"></span>Ｂ<span class="_ _6"></span>Ｎ</div><div class="t m0 x5 hd yf8 ff8 fs5 fc2 sc2 ls0 ws0">来<span class="_ _f"></span>初<span class="_ _f"></span>始<span class="_ _e"></span>化<span class="_ _f"></span>的</div><div class="t m0 xa3 h11 yf9 ff2 fs5 fc2 sc2 ls0 ws0">Ｍ<span class="_ _c"></span>Ｌ<span class="_ _2"></span>Ｐ</div><div class="t m0 x92 hd yf8 ff8 fs5 fc2 sc2 ls0 ws0">为</div><div class="t m0 x17c h11 yf9 ff2 fs5 fc2 sc2 ls0 ws0">Ｄ<span class="_ _3"></span>Ｎ<span class="_ _c"></span>Ｎ</div><div class="t m0 x31 h16 yfa ffd fs5 fc2 sc2 ls0 ws0">。</div><div class="t m0 x145 h11 yf9 ff2 fs5 fc2 sc2 ls0 ws0">Ｄ<span class="_ _3"></span>Ｎ<span class="_ _c"></span>Ｎ</div><div class="t m0 x17d hd yf8 ff8 fs5 fc2 sc2 ls9 ws0">模型的训练阶段大致分为两个步骤</div><div class="t m0 xd7 h16 yfa ffd fs5 fc2 sc2 ls0 ws0">：<span class="_ _1"></span>（</div><div class="t m0 xbb h11 yf9 ff2 fs5 fc2 sc2 ls0 ws0">１</div><div class="t m0 x17e h16 yfa ffd fs5 fc2 sc2 ls0 ws0">）</div><div class="t m0 x174 hd yf8 ff8 fs5 fc2 sc2 ls0 ws0">预<span class="_ _17"> </span>训<span class="_ _17"> </span>练</div><div class="t m0 x3a h16 yfa ffd fs5 fc2 sc2 ls0 ws0">（</div><div class="t m0 x162 h11 yf9 ff2 fs5 fc2 sc2 ls0 ws0">Ｐ<span class="_ _7"></span>ｒ<span class="_ _a"></span>ｅ</div><div class="t m4 x17f h11 yfb ff2 fs5 fc2 sc2 ls0 ws0">－</div><div class="t m0 x180 h11 yf9 ff2 fs5 fc2 sc2 ls0 ws0">ｔ<span class="_ _b"></span>ｒ<span class="_ _a"></span>ａ<span class="_ _13"></span>ｉ<span class="_ _a"></span>ｎ</div><div class="t m4 xeb h11 yfb ff2 fs5 fc2 sc2 ls0 ws0">－</div><div class="t m0 x159 h11 yfc ff2 fs5 fc2 sc2 ls0 ws0">ｉ<span class="_ _b"></span>ｎ</div><div class="t m0 x15a h11 yfd ff2 fs5 fc2 sc2 ls0 ws0">ｇ</div><div class="t m0 x84 h16 yfe ffd fs5 fc2 sc2 ls0 ws0">）<span class="_ _8"></span>，</div><div class="t m0 xb hd yff ff8 fs5 fc2 sc2 ls4 ws0">利用无监督学习的算法来训练受限波尔兹曼机</div><div class="t m0 x15b h16 yfe ffd fs5 fc2 sc2 ls0 ws0">（</div><div class="t m0 x181 h11 yfc ff2 fs5 fc2 sc2 ls0 ws0">Ｒ<span class="_ _1"></span>ｅ<span class="_ _9"></span>ｓ<span class="_ _b"></span>ｔ<span class="_ _b"></span>ｒ<span class="_ _13"></span>ｉ<span class="_ _13"></span>ｃ<span class="_ _b"></span>ｔ<span class="_ _b"></span>ｅ<span class="_ _7"></span>ｄ</div><div class="t m0 x5a he y100 ff3 fs5 fc2 sc2 ls0 ws0">　</div><div class="t m0 x55 h11 yfc ff2 fs5 fc2 sc2 ls0 ws0">Ｂ<span class="_ _1"></span>ｏ<span class="_ _b"></span>ｌ<span class="_ _13"></span>ｔ<span class="_ _b"></span>ｚ<span class="_ _4"></span>ｍ<span class="_ _2"></span>ａ<span class="_ _8"></span>ｎ<span class="_ _8"></span>ｎ</div><div class="t m0 x16a he y100 ff3 fs5 fc2 sc2 ls0 ws0">　</div><div class="t m0 xb7 h11 yfc ff2 fs5 fc2 sc2 ls0 ws0">ｍ<span class="_ _2"></span>ａ<span class="_ _9"></span>ｃ<span class="_ _7"></span>ｈ<span class="_ _b"></span>ｉ<span class="_ _b"></span>ｎ<span class="_ _7"></span>ｅ</div><div class="t m0 x61 h16 yfe ffd fs5 fc2 sc2 ls0 ws0">，</div><div class="t m0 x174 h11 yfc ff2 fs5 fc2 sc2 ls0 ws0">Ｒ<span class="_ _6"></span>Ｂ<span class="_ _5"></span>Ｍ</div><div class="t m0 x42 h16 yfe ffd fs5 fc2 sc2 ls0 ws0">）<span class="_ _8"></span>，</div><div class="t m0 x146 h11 yfc ff2 fs5 fc2 sc2 ls0 ws0">Ｒ<span class="_ _6"></span>Ｂ<span class="_ _5"></span>Ｍ</div><div class="t m0 x149 hd yff ff8 fs5 fc2 sc2 ls0 ws0">通<span class="_ _e"></span>过</div><div class="t m0 x27 hd y101 ff8 fs5 fc2 sc2 ls4 ws0">逐层训练并堆叠成</div><div class="t m0 x7c h11 y102 ff2 fs5 fc2 sc2 ls0 ws0">Ｄ<span class="_ _6"></span>Ｂ<span class="_ _6"></span>Ｎ</div><div class="t m0 x182 h16 y103 ffd fs5 fc2 sc2 ls0 ws0">；<span class="_ _8"></span>（</div><div class="t m0 xa4 h11 y102 ff2 fs5 fc2 sc2 ls0 ws0">２</div><div class="t m0 xfd h16 y103 ffd fs5 fc2 sc2 ls0 ws0">）</div><div class="t m0 x94 hd y101 ff8 fs5 fc2 sc2 ls0 ws0">模<span class="_ _17"></span>型<span class="_ _17"></span>精<span class="_ _17"></span>细<span class="_ _e"></span>调<span class="_ _17"> </span>整</div><div class="t m0 x183 h16 y103 ffd fs5 fc2 sc2 ls0 ws0">（</div><div class="t m0 x77 h11 y102 ff2 fs5 fc2 sc2 ls0 ws0">Ｆ<span class="_ _a"></span>ｉ<span class="_ _a"></span>ｎ<span class="_ _7"></span>ｅ</div><div class="t m4 x5d h11 y104 ff2 fs5 fc2 sc2 ls0 ws0">－</div><div class="t m0 x16d h11 y102 ff2 fs5 fc2 sc2 ls0 ws0">ｔ<span class="_ _9"></span>ｕ<span class="_ _8"></span>ｎ<span class="_ _b"></span>ｉ<span class="_ _b"></span>ｎ</div><div class="t m0 x184 h11 y105 ff2 fs5 fc2 sc2 ls0 ws0">ｇ</div><div class="t m0 x139 h16 y103 ffd fs5 fc2 sc2 ls0 ws0">）<span class="_ _8"></span>，</div><div class="t m0 x15d hd y101 ff8 fs5 fc2 sc2 ls0 ws0">在</div><div class="t m0 xf5 h11 y102 ff2 fs5 fc2 sc2 ls0 ws0">Ｄ<span class="_ _6"></span>Ｂ<span class="_ _6"></span>Ｎ</div><div class="t m0 x170 hd y101 ff8 fs5 fc2 sc2 lsa ws0">的最后一层上面增加一层</div><div class="t m0 xe9 h11 y102 ff2 fs5 fc2 sc2 ls0 ws0">Ｓ<span class="_ _8"></span>ｏ<span class="_ _b"></span>ｆ<span class="_ _b"></span>ｔ<span class="_ _1"></span>ｍ<span class="_ _6"></span>ａ<span class="_ _7"></span>ｘ</div><div class="t m0 x27 hd y106 ff8 fs5 fc2 sc2 ls0 ws0">层</div><div class="t m0 x56 h16 y107 ffd fs5 fc2 sc2 ls0 ws0">，</div><div class="t m0 x3 hd y106 ff8 fs5 fc2 sc2 ls4 ws0">将其用于初始化</div><div class="t m0 x185 h11 y108 ff2 fs5 fc2 sc2 ls0 ws0">Ｄ<span class="_ _3"></span>Ｎ<span class="_ _c"></span>Ｎ</div><div class="t m0 x14f hd y106 ff8 fs5 fc2 sc2 ls0 ws0">的<span class="_ _f"></span>模<span class="_ _f"></span>型<span class="_ _e"></span>参<span class="_ _f"></span>数</div><div class="t m0 x112 h16 y107 ffd fs5 fc2 sc2 ls0 ws0">，</div><div class="t m0 x186 hd y106 ff8 fs5 fc2 sc2 ls0 ws0">然<span class="_ _e"></span>后<span class="_ _e"></span>使<span class="_ _17"></span>用<span class="_ _e"></span>带<span class="_ _17"></span>标<span class="_ _17"></span>注<span class="_ _e"></span>的<span class="_ _17"></span>数<span class="_ _e"></span>据</div><div class="t m0 x5f h16 y107 ffd fs5 fc2 sc2 ls0 ws0">，</div><div class="t m0 xe0 hd y106 ff8 fs5 fc2 sc2 ls8 ws0">利用传统神经网络的学习算法</div><div class="t m0 x148 h16 y107 ffd fs5 fc2 sc2 ls0 ws0">（</div><div class="t m0 x187 hd y106 ff8 fs5 fc2 sc2 ls0 ws0">如</div><div class="t m0 xc0 h11 y108 ff2 fs5 fc2 sc2 ls0 ws0">Ｂ<span class="_ _4"></span>Ｐ</div><div class="t m0 x27 hd y109 ff8 fs5 fc2 sc2 ls0 ws0">算<span class="_ _f"></span>法</div><div class="t m0 x57 h16 y10a ffd fs5 fc2 sc2 ls0 ws0">）</div><div class="t m0 xa0 hd y109 ff8 fs5 fc2 sc2 ls0 ws0">来<span class="_ _f"></span>学<span class="_ _f"></span>习</div><div class="t m0 x24 h11 y10b ff2 fs5 fc2 sc2 ls0 ws0">Ｄ<span class="_ _3"></span>Ｎ<span class="_ _c"></span>Ｎ</div><div class="t m0 x7c hd y109 ff8 fs5 fc2 sc2 ls0 ws0">的<span class="_ _f"></span>模<span class="_ _f"></span>型<span class="_ _e"></span>参<span class="_ _f"></span>数</div><div class="t m0 x74 h16 y10a ffd fs5 fc2 sc2 ls0 ws0">。</div><div class="t m0 x188 hd y109 ff8 fs5 fc2 sc2 ls0 ws0">如<span class="_ _f"></span>此</div><div class="t m0 x112 h16 y10a ffd fs5 fc2 sc2 ls0 ws0">，</div><div class="t m0 xa8 hd y109 ff8 fs5 fc2 sc2 ls4 ws0">具有很多隐层的</div><div class="t m0 x155 h16 y10a ffd fs5 fc2 sc2 ls0 ws0">（</div><div class="t m0 x53 hd y109 ff8 fs5 fc2 sc2 ls0 ws0">即<span class="_ _f"></span>深<span class="_ _f"></span>层<span class="_ _e"></span>的</div><div class="t m0 x104 h16 y10a ffd fs5 fc2 sc2 ls0 ws0">，</div><div class="t m0 x189 hd y109 ff8 fs5 fc2 sc2 ls4 ws0">一般指隐层数大于</div><div class="t m0 x34 h11 y10b ff2 fs5 fc2 sc2 ls0 ws0">２</div><div class="t m0 x10b hd y109 ff8 fs5 fc2 sc2 ls0 ws0">乃<span class="_ _f"></span>至<span class="_ _f"></span>几<span class="_ _e"></span>百<span class="_ _f"></span>上<span class="_ _f"></span>千</div><div class="t m0 xeb h16 y10a ffd fs5 fc2 sc2 ls0 ws0">）</div><div class="t m0 x27 hd y10c ff8 fs5 fc2 sc2 ls4 ws0">的大规模模型参数</div><div class="t m0 xc4 h16 y10d ffd fs5 fc2 sc2 ls0 ws0">（</div><div class="t m0 x9 hd y10c ff8 fs5 fc2 sc2 ls4 ws0">一般参数数量百万级左右或以上</div><div class="t m0 x63 h16 y10d ffd fs5 fc2 sc2 ls0 ws0">）</div><div class="t m0 x64 hd y10c ff8 fs5 fc2 sc2 ls4 ws0">的学习或训练问题在训练数据充分的条件下<span class="_ _10"></span>一定</div><div class="t m0 x27 hd y10e ff8 fs5 fc2 sc2 ls4 ws0">程度上得到了解决</div><div class="t m0 x5c h16 y10f ffd fs5 fc2 sc2 ls0 ws0">，</div><div class="t m0 x9 hd y10e ff8 fs5 fc2 sc2 ls4 ws0">使得其强大的学习和表达能力在机器学习中得以发挥</div><div class="t m0 x9d h16 y10f ffd fs5 fc2 sc2 ls0 ws0">，</div><div class="t m0 xb7 hd y10e ff8 fs5 fc2 sc2 ls4 ws0">也直接导致机器学习领<span class="_ _10"></span>域掀</div><div class="t m0 x27 hd y110 ff8 fs5 fc2 sc2 ls4 ws0">起了深度学习的热潮</div><div class="t m0 x12 h16 y111 ffd fs5 fc2 sc2 ls0 ws0">，</div><div class="t m0 x18a hd y110 ff8 fs5 fc2 sc2 ls0 ws0">同<span class="_ _f"></span>时</div><div class="t m0 x88 h16 y111 ffd fs5 fc2 sc2 ls0 ws0">，</div><div class="t m0 x66 hd y110 ff8 fs5 fc2 sc2 ls0 ws0">有<span class="_ _f"></span>别<span class="_ _f"></span>于</div><div class="t m0 x8b h11 y112 ff2 fs5 fc2 sc2 ls0 ws0">Ｍ<span class="_ _c"></span>Ｌ<span class="_ _2"></span>Ｐ</div><div class="t m0 xcb hd y110 ff8 fs5 fc2 sc2 ls4 ws0">的各种新的深层神经网络结构模型也被提出</div><div class="t m0 x18b h16 y111 ffd fs5 fc2 sc2 ls0 ws0">。</div><div class="t m0 x56 h1d y113 ff2 fs8 fc2 sc2 ls0 ws0">２<span class="_ _18"></span>２<span class="_ _18"></span>２</div><div class="t m0 x18c hd y114 ffa fs5 fc2 sc2 ls4 ws0">数据采集与处理</div><div class="t m6 x112 h1e y115 ffe fs9 fc2 sc2 ls0 ws0">Ｊ<span class="_ _8"></span>ｏ<span class="_ _7"></span>ｕ<span class="_ _9"></span>ｒ<span class="_ _7"></span>ｎ<span class="_ _7"></span>ａ<span class="_ _a"></span>ｌ</div><div class="t m0 x82 he y116 ff3 fs5 fc2 sc2 ls0 ws0">　</div><div class="t m6 x99 h1e y115 ffe fs9 fc2 sc2 ls0 ws0">ｏ</div><div class="t m6 xf1 h1e y117 ffe fs9 fc2 sc2 ls0 ws0">ｆ</div><div class="t m0 x5d he y118 ff3 fs5 fc2 sc2 ls0 ws0">　</div><div class="t m6 x16d h1e y115 ffe fs9 fc2 sc2 ls0 ws0">Ｄ<span class="_ _2"></span>ａ<span class="_ _a"></span>ｔ<span class="_ _b"></span>ａ</div><div class="t m0 x184 he y116 ff3 fs5 fc2 sc2 ls0 ws0">　</div><div class="t m6 x155 h1e y115 ffe fs9 fc2 sc2 ls0 ws0">Ａ<span class="_ _1"></span>ｃ</div><div class="t m6 x151 h1e y117 ffe fs9 fc2 sc2 ls0 ws0">ｑ</div><div class="t m6 x13c h1e y115 ffe fs9 fc2 sc2 ls0 ws0">ｕ<span class="_ _a"></span>ｉ<span class="_ _13"></span>ｓ<span class="_ _13"></span>ｉ<span class="_ _13"></span>ｔ<span class="_ _13"></span>ｉ<span class="_ _b"></span>ｏ<span class="_ _9"></span>ｎ</div><div class="t m0 x18d he y116 ff3 fs5 fc2 sc2 ls0 ws0">　</div><div class="t m6 x9c h1e y115 ffe fs9 fc2 sc2 ls0 ws0">ａ<span class="_ _7"></span>ｎ<span class="_ _1"></span>ｄ</div><div class="t m0 x18e he y116 ff3 fs5 fc2 sc2 ls0 ws0">　</div><div class="t m6 xe1 h1e y115 ffe fs9 fc2 sc2 ls0 ws0">Ｐ<span class="_ _1"></span>ｒ<span class="_ _9"></span>ｏ<span class="_ _a"></span>ｃ<span class="_ _b"></span>ｅ<span class="_ _b"></span>ｓ<span class="_ _b"></span>ｓ<span class="_ _13"></span>ｉ<span class="_ _b"></span>ｎ</div><div class="t m6 x45 h1e y117 ffe fs9 fc2 sc2 ls0 ws0">ｇ</div><div class="t m0 x18f he y118 ff3 fs5 fc2 sc2 ls0 ws0">　</div><div class="t m0 x61 h11 y115 ff2 fs5 fc2 sc2 ls0 ws0">Ｖ<span class="_ _2"></span>ｏ<span class="_ _b"></span>ｌ<span class="_ _4"></span>．<span class="_ _11"></span>３<span class="_ _8"></span>２</div><div class="t m0 x42 h1f y119 fff fs5 fc2 sc2 ls0 ws0">，</div><div class="t m0 x39 h11 y115 ff2 fs5 fc2 sc2 ls0 ws0">Ｎ<span class="_ _2"></span>ｏ<span class="_ _3"></span>．<span class="_ _15"></span>２</div><div class="t m0 x17f h1f y119 fff fs5 fc2 sc2 ls0 ws0">，</div><div class="t m0 x190 h11 y115 ff2 fs5 fc2 sc2 ls0 ws0">２<span class="_ _7"></span>０<span class="_ _8"></span>１<span class="_ _8"></span>７</div></div><div class="pi" data-data='{"ctm":[1.830664,0.000000,0.000000,1.830664,0.000000,0.000000]}'></div></div>
<div id="pf3" class="pf w0 h0" data-page-no="3"><div class="pc pc3 w0 h0"><img class="bi x0 y0 w1 h1" alt="" src="https://static.pudn.com/prod/directory_preview_static/625fb7df8cbeb85d57682fe6/bg3.jpg"><div class="t m0 x57 hd y8f ff8 fs5 fc2 sc2 ls0 ws0">早<span class="_ _f"></span>期<span class="_ _f"></span>的</div><div class="t m0 xe h11 y11a ff2 fs5 fc2 sc2 ls0 ws0">Ｄ<span class="_ _3"></span>Ｎ<span class="_ _c"></span>Ｎ</div><div class="t m0 x90 hd y8f ff8 fs5 fc2 sc2 lsa ws0">主<span class="_ _16"></span>要<span class="_ _d"></span>是<span class="_ _d"></span>特指前馈全连接深层神经网络</div><div class="t m0 x1f h16 y90 ff10 fs5 fc2 sc2 ls0 ws0">（</div><div class="t m0 x191 h11 y11a ff2 fs5 fc2 sc2 ls0 ws0">Ｆ<span class="_ _8"></span>ｅ<span class="_ _9"></span>ｅ<span class="_ _7"></span>ｄ<span class="_ _a"></span>ｆ<span class="_ _a"></span>ｏ<span class="_ _a"></span>ｒ<span class="_ _4"></span>ｗ<span class="_ _4"></span>ａ<span class="_ _a"></span>ｒ<span class="_ _9"></span>ｄ</div><div class="t m0 x192 he y11b ff3 fs5 fc2 sc2 ls0 ws0">　</div><div class="t m0 x119 h11 y11a ff2 fs5 fc2 sc2 ls0 ws0">ｆ<span class="_ _a"></span>ｕ<span class="_ _a"></span>ｌ<span class="_ _12"></span>ｌ</div><div class="t m0 x193 h11 y11c ff2 fs5 fc2 sc2 ls0 ws0">ｙ</div><div class="t m4 x131 h11 y11d ff2 fs5 fc2 sc2 ls0 ws0">－</div><div class="t m0 xb8 h11 y11a ff2 fs5 fc2 sc2 ls0 ws0">ｃ<span class="_ _9"></span>ｏ<span class="_ _7"></span>ｎ<span class="_ _1"></span>ｎ<span class="_ _9"></span>ｅ<span class="_ _9"></span>ｃ<span class="_ _b"></span>ｔ<span class="_ _b"></span>ｅ<span class="_ _7"></span>ｄ</div><div class="t m0 x34 he y11b ff3 fs5 fc2 sc2 ls0 ws0">　</div><div class="t m0 xfa h11 y11a ff2 fs5 fc2 sc2 ls0 ws0">ｄ<span class="_ _9"></span>ｅ<span class="_ _9"></span>ｅ</div><div class="t m0 x41 h11 y11c ff2 fs5 fc2 sc2 ls0 ws0">ｐ</div><div class="t m0 xbd he y11e ff3 fs5 fc2 sc2 ls0 ws0">　</div><div class="t m0 xbe h11 y11a ff2 fs5 fc2 sc2 ls0 ws0">ｎ<span class="_ _7"></span>ｅ<span class="_ _7"></span>ｕ<span class="_ _9"></span>ｒ<span class="_ _a"></span>ａ<span class="_ _b"></span>ｌ</div><div class="t m0 x190 he y11b ff3 fs5 fc2 sc2 ls0 ws0">　</div><div class="t m0 xea h11 y11a ff2 fs5 fc2 sc2 ls0 ws0">ｎ<span class="_ _7"></span>ｅ<span class="_ _b"></span>ｔ</div><div class="t m4 xeb h11 y11d ff2 fs5 fc2 sc2 ls0 ws0">－</div><div class="t m0 x20 h11 y11f ff2 fs5 fc2 sc2 ls0 ws0">ｗ<span class="_ _4"></span>ｏ<span class="_ _9"></span>ｒ<span class="_ _9"></span>ｋ<span class="_ _7"></span>ｓ</div><div class="t m0 xc h16 y120 ff10 fs5 fc2 sc2 ls0 ws0">，</div><div class="t m0 x4c h11 y11f ff2 fs5 fc2 sc2 ls0 ws0">Ｆ<span class="_ _3"></span>Ｎ<span class="_ _5"></span>Ｎ</div><div class="t m0 x7 h16 y120 ff10 fs5 fc2 sc2 ls0 ws0">）<span class="_ _8"></span>。</div><div class="t m0 x70 hd y121 ff8 fs5 fc2 sc2 ls4 ws0">此后随着深度学习的发展</div><div class="t m0 x194 h16 y120 ff10 fs5 fc2 sc2 ls0 ws0">，</div><div class="t m0 x195 hd y121 ff8 fs5 fc2 sc2 ls0 ws0">卷<span class="_ _f"></span>积<span class="_ _f"></span>神<span class="_ _e"></span>经<span class="_ _f"></span>网<span class="_ _f"></span>络</div><div class="t m0 x54 h16 y120 ff10 fs5 fc2 sc2 ls0 ws0">（</div><div class="t m0 x196 h11 y11f ff2 fs5 fc2 sc2 ls0 ws0">Ｃ<span class="_ _1"></span>ｏ<span class="_ _8"></span>ｎ<span class="_ _7"></span>ｖ<span class="_ _7"></span>ｏ<span class="_ _b"></span>ｌ<span class="_ _a"></span>ｕ<span class="_ _9"></span>ｔ<span class="_ _12"></span>ｉ<span class="_ _b"></span>ｏ<span class="_ _7"></span>ｎ<span class="_ _7"></span>ａ<span class="_ _b"></span>ｌ</div><div class="t m0 x152 he y122 ff3 fs5 fc2 sc2 ls0 ws0">　</div><div class="t m0 x105 h11 y11f ff2 fs5 fc2 sc2 ls0 ws0">ｎ<span class="_ _7"></span>ｅ<span class="_ _7"></span>ｕ<span class="_ _9"></span>ｒ<span class="_ _a"></span>ａ<span class="_ _b"></span>ｌ</div><div class="t m0 x197 he y122 ff3 fs5 fc2 sc2 ls0 ws0">　</div><div class="t m0 x198 h11 y11f ff2 fs5 fc2 sc2 ls0 ws0">ｎ<span class="_ _7"></span>ｅ<span class="_ _b"></span>ｔ<span class="_ _1"></span>ｗ<span class="_ _2"></span>ｏ<span class="_ _a"></span>ｒ<span class="_ _7"></span>ｋ<span class="_ _7"></span>ｓ</div><div class="t m0 x199 h16 y120 ff10 fs5 fc2 sc2 ls0 ws0">，</div><div class="t m0 x41 h11 y11f ff2 fs5 fc2 sc2 ls0 ws0">Ｃ<span class="_ _3"></span>Ｎ<span class="_ _5"></span>Ｎ</div><div class="t m0 x49 h16 y120 ff10 fs5 fc2 sc2 ls0 ws0">）</div><div class="t m0 x148 hd y121 ff8 fs5 fc2 sc2 ls0 ws0">和<span class="_ _f"></span>递<span class="_ _f"></span>归</div><div class="t m0 x27 hd y123 ff8 fs5 fc2 sc2 ls0 ws0">神<span class="_ _f"></span>经<span class="_ _f"></span>网<span class="_ _e"></span>络</div><div class="t m0 xaa h16 y124 ff10 fs5 fc2 sc2 ls0 ws0">（</div><div class="t m0 x110 h11 y125 ff2 fs5 fc2 sc2 ls0 ws0">Ｒ<span class="_ _1"></span>ｅ<span class="_ _9"></span>ｃ<span class="_ _9"></span>ｕ<span class="_ _9"></span>ｒ<span class="_ _a"></span>ｒ<span class="_ _a"></span>ｅ<span class="_ _7"></span>ｎ<span class="_ _9"></span>ｔ</div><div class="t m0 x4f he y126 ff3 fs5 fc2 sc2 ls0 ws0">　</div><div class="t m0 xa h11 y125 ff2 fs5 fc2 sc2 ls0 ws0">ｎ<span class="_ _7"></span>ｅ<span class="_ _7"></span>ｕ<span class="_ _9"></span>ｒ<span class="_ _a"></span>ａ<span class="_ _b"></span>ｌ</div><div class="t m0 x19a he y126 ff3 fs5 fc2 sc2 ls0 ws0">　</div><div class="t m0 xdb h11 y125 ff2 fs5 fc2 sc2 ls0 ws0">ｎ<span class="_ _7"></span>ｅ<span class="_ _b"></span>ｔ<span class="_ _1"></span>ｗ<span class="_ _2"></span>ｏ<span class="_ _a"></span>ｒ<span class="_ _7"></span>ｋ<span class="_ _7"></span>ｓ</div><div class="t m0 xb0 h16 y124 ff10 fs5 fc2 sc2 ls0 ws0">，</div><div class="t m0 x19b h11 y125 ff2 fs5 fc2 sc2 ls0 ws0">Ｒ<span class="_ _3"></span>Ｎ<span class="_ _c"></span>Ｎ</div><div class="t m0 x183 h16 y124 ff10 fs5 fc2 sc2 ls0 ws0">）</div><div class="t m0 x169 hd y123 ff8 fs5 fc2 sc2 ls4 ws0">等网络结构在<span class="ls5">机器学习不同任务中得到应用</span></div><div class="t m0 x19c h16 y124 ff10 fs5 fc2 sc2 ls0 ws0">，</div><div class="t m0 x43 hd y123 ff8 fs5 fc2 sc2 ls0 ws0">并<span class="_ _e"></span>且<span class="_ _e"></span>相<span class="_ _e"></span>比<span class="_ _e"></span>于</div><div class="t m0 x6e h11 y127 ff2 fs5 fc2 sc2 ls0 ws0">Ｄ<span class="_ _3"></span>Ｎ<span class="_ _c"></span>Ｎ</div><div class="t m0 x11f hd y128 ff8 fs5 fc2 sc2 ls4 ws0">展现出各自的优势</div><div class="t m0 x121 h16 y129 ff10 fs5 fc2 sc2 ls0 ws0">，</div><div class="t m0 x91 hd y128 ff8 fs5 fc2 sc2 ls4 ws0">受到越来越广泛的关注</div><div class="t m0 x19d h16 y129 ff10 fs5 fc2 sc2 ls0 ws0">。</div><div class="t m0 x159 h12 y12a ffc fs6 fc2 sc2 ls0 ws0">２</div><div class="t m0 x15a h13 y12b ff3 fs6 fc2 sc2 ls0 ws0">　</div><div class="t m0 x4 h18 y12c ff6 fs6 fc2 sc2 lsb ws0">基于深度学习的语音识别技术</div><div class="t m0 x159 h20 y12d ffc fs8 fc2 sc2 ls0 ws0">２<span class="_ _5"></span>．<span class="_ _14"></span>１</div><div class="t m0 x3 h21 y12e ff3 fs8 fc2 sc2 ls0 ws0">　</div><div class="t m0 x6f h22 y12f ff6 fs8 fc2 sc2 ls0 ws0">基<span class="_ _f"></span>于<span class="_ _f"></span>深<span class="_ _f"></span>度<span class="_ _f"></span>学<span class="_ _f"></span>习<span class="_ _f"></span>的<span class="_ _f"></span>声<span class="_ _f"></span>学<span class="_ _f"></span>模<span class="_ _f"></span>型<span class="_ _f"></span>训<span class="_ _f"></span>练<span class="_ _f"></span>准<span class="_ _f"></span>则</div><div class="t m0 x27 he y130 ff3 fs5 fc2 sc2 ls0 ws0">　<span class="_ _f"></span>　</div><div class="t m0 x57 hd y131 ff8 fs5 fc2 sc2 ls0 ws0">本<span class="_ _f"></span>节<span class="_ _f"></span>以<span class="_ _e"></span>基<span class="_ _f"></span>于</div><div class="t m0 x87 h11 y132 ff2 fs5 fc2 sc2 ls0 ws0">Ｄ<span class="_ _3"></span>Ｎ<span class="_ _c"></span>Ｎ</div><div class="t m4 xee h11 y130 ff2 fs5 fc2 sc2 ls0 ws0">－</div><div class="t m0 x6b h11 y132 ff2 fs5 fc2 sc2 ls0 ws0">Ｈ<span class="_ _d"></span>ＭＭ</div><div class="t m0 x12b hd y131 ff8 fs5 fc2 sc2 ls8 ws0">的<span class="_ _d"></span>语<span class="_ _d"></span>音<span class="_ _d"></span>识<span class="_ _16"></span>别声学模<span class="_ _d"></span>型框架讨论基于深度学习的声学模型训练准则</div><div class="t m0 x14b h16 y133 ff10 fs5 fc2 sc2 ls0 ws0">，</div><div class="t m0 x149 hd y131 ff8 fs5 fc2 sc2 ls0 ws0">该<span class="_ _e"></span>框</div><div class="t m0 x27 hd y134 ff8 fs5 fc2 sc2 ls0 ws0">架<span class="_ _f"></span>如<span class="_ _f"></span>图</div><div class="t m0 x29 h11 y135 ff2 fs5 fc2 sc2 ls0 ws0">１</div><div class="t m0 x19e hd y134 ff8 fs5 fc2 sc2 ls0 ws0">所<span class="_ _17"> </span>示</div><div class="t m0 x10 h16 y136 ff10 fs5 fc2 sc2 ls0 ws0">。</div><div class="t m0 x2b hd y134 ff8 fs5 fc2 sc2 ls9 ws0">相比于传统的基于</div><div class="t m0 x124 h11 y135 ff2 fs5 fc2 sc2 ls0 ws0">Ｇ<span class="_ _0"></span>ＭＭ</div><div class="t m4 x195 h11 y137 ff2 fs5 fc2 sc2 ls0 ws0">－</div><div class="t m0 x183 h11 y135 ff2 fs5 fc2 sc2 ls0 ws0">Ｈ<span class="_ _d"></span>ＭＭ</div><div class="t m0 xde hd y134 ff8 fs5 fc2 sc2 ls9 ws0">的语音识别框架</div><div class="t m0 xd2 h16 y136 ff10 fs5 fc2 sc2 ls0 ws0">，</div><div class="t m0 x118 hd y134 ff8 fs5 fc2 sc2 ls9 ws0">其最大的改变是采用</div><div class="t m0 x3a h11 y135 ff2 fs5 fc2 sc2 ls0 ws0">Ｄ<span class="_ _3"></span>Ｎ<span class="_ _c"></span>Ｎ</div><div class="t m0 x149 hd y134 ff8 fs5 fc2 sc2 ls0 ws0">替<span class="_ _17"> </span>换</div><div class="t m0 x20 h11 y138 ff2 fs5 fc2 sc2 ls0 ws0">Ｇ<span class="_ _0"></span>ＭＭ</div><div class="t m0 x19f hd y139 ff8 fs5 fc2 sc2 ls4 ws0">模型来对语音的观察概率进行建模</div><div class="t m0 x32 h16 y13a ff10 fs5 fc2 sc2 ls0 ws0">。</div><div class="t m0 x113 h11 y138 ff2 fs5 fc2 sc2 ls0 ws0">Ｄ<span class="_ _3"></span>Ｎ<span class="_ _c"></span>Ｎ</div><div class="t m0 x16d hd y139 ff8 fs5 fc2 sc2 ls0 ws0">相<span class="_ _e"></span>比<span class="_ _e"></span>于</div><div class="t m0 x54 h11 y138 ff2 fs5 fc2 sc2 ls0 ws0">Ｇ<span class="_ _0"></span>ＭＭ</div><div class="t m0 x55 hd y139 ff8 fs5 fc2 sc2 ls0 ws0">的<span class="_ _e"></span>优<span class="_ _e"></span>势<span class="_ _e"></span>在<span class="_ _e"></span>于</div><div class="t m0 x1a0 h16 y13a ff10 fs5 fc2 sc2 ls0 ws0">：<span class="_ _7"></span>（</div><div class="t m0 xf8 h11 y138 ff2 fs5 fc2 sc2 ls0 ws0">１</div><div class="t m0 x1a1 h16 y13a ff10 fs5 fc2 sc2 ls0 ws0">）</div><div class="t m0 x167 hd y139 ff8 fs5 fc2 sc2 ls0 ws0">使<span class="_ _e"></span>用</div><div class="t m0 x34 h11 y138 ff2 fs5 fc2 sc2 ls0 ws0">Ｄ<span class="_ _3"></span>Ｎ<span class="_ _c"></span>Ｎ</div><div class="t m0 xbe hd y139 ff8 fs5 fc2 sc2 ls0 ws0">估<span class="_ _e"></span>计</div><div class="t m0 x180 h11 y138 ff2 fs5 fc2 sc2 ls0 ws0">Ｈ<span class="_ _d"></span>ＭＭ</div><div class="t m0 x27 hd y13b ff8 fs5 fc2 sc2 ls4 ws0">状态的后验概率分布不需要对语音数据分布进行假设</div><div class="t m0 x64 h16 y13c ff10 fs5 fc2 sc2 ls0 ws0">；<span class="_ _9"></span>（</div><div class="t m0 x79 h11 y13d ff2 fs5 fc2 sc2 ls0 ws0">２</div><div class="t m0 x156 h16 y13c ff10 fs5 fc2 sc2 ls0 ws0">）</div><div class="t m0 x1a2 h11 y13d ff2 fs5 fc2 sc2 ls0 ws0">Ｄ<span class="_ _3"></span>Ｎ<span class="_ _c"></span>Ｎ</div><div class="t m0 x1a3 hd y13b ff8 fs5 fc2 sc2 ls4 ws0">的输入特征可以<span class="ls5">是多种特征的融合</span></div><div class="t m0 xda h16 y13c ff10 fs5 fc2 sc2 ls0 ws0">，</div><div class="t m0 xc1 hd y13b ff8 fs5 fc2 sc2 ls0 ws0">包</div><div class="t m0 x27 hd y13e ff8 fs5 fc2 sc2 ls4 ws0">括离散或者连续的</div><div class="t m0 x5c h16 y13f ff10 fs5 fc2 sc2 ls0 ws0">；<span class="_ _9"></span>（</div><div class="t m0 xa h11 y140 ff2 fs5 fc2 sc2 ls0 ws0">３</div><div class="t m0 xee h16 y13f ff10 fs5 fc2 sc2 ls0 ws0">）</div><div class="t m0 x72 h11 y140 ff2 fs5 fc2 sc2 ls0 ws0">Ｄ<span class="_ _3"></span>Ｎ<span class="_ _c"></span>Ｎ</div><div class="t m0 x17c hd y13e ff8 fs5 fc2 sc2 ls4 ws0">可以利用相邻语音帧所包含的结构信息</div><div class="t m0 xd3 h16 y13f ff10 fs5 fc2 sc2 ls0 ws0">。</div><div class="t m0 x138 h23 y141 ff8 fsa fc2 sc2 ls0 ws0">图</div><div class="t m0 x1a3 h24 y142 ff2 fsa fc2 sc2 ls0 ws0">１</div><div class="t m0 x104 h25 y143 ff3 fsa fc2 sc2 ls0 ws0">　</div><div class="t m0 x12f h23 y141 ff8 fsa fc2 sc2 ls0 ws0">基<span class="_ _f"></span>于</div><div class="t m0 xe2 h24 y142 ff2 fsa fc2 sc2 ls0 ws0">Ｄ<span class="_ _5"></span>Ｎ<span class="_ _c"></span>Ｎ</div><div class="t m7 x107 h24 y143 ff2 fsa fc2 sc2 ls0 ws0">－</div><div class="t m0 x197 h24 y142 ff2 fsa fc2 sc2 ls0 ws0">Ｈ<span class="_ _d"></span>ＭＭ</div><div class="t m0 x1a4 h23 y141 ff8 fsa fc2 sc2 lsc ws0">的语音识别系统</div><div class="t m0 xd0 h25 y144 ff3 fsa fc2 sc2 ls0 ws0">　</div><div class="t m0 x116 h24 y145 ff2 fsa fc2 sc2 ls0 ws0">Ｆ<span class="_ _7"></span>ｉ</div><div class="t m0 x16f h24 y146 ff2 fsa fc2 sc2 ls0 ws0">ｇ</div><div class="t m0 x1a3 h24 y145 ff2 fsa fc2 sc2 ls0 ws0">．<span class="_ _13"></span>１</div><div class="t m0 x18d h25 y144 ff3 fsa fc2 sc2 ls0 ws0">　</div><div class="t m0 x1a5 h24 y145 ff2 fsa fc2 sc2 ls0 ws0">Ｉ<span class="_ _b"></span>ｌ<span class="_ _13"></span>ｌ<span class="_ _9"></span>ｕ<span class="_ _8"></span>ｓ<span class="_ _9"></span>ｔ<span class="_ _a"></span>ｒ<span class="_ _9"></span>ａ<span class="_ _9"></span>ｔ<span class="_ _b"></span>ｉ<span class="_ _a"></span>ｏ<span class="_ _1"></span>ｎ</div><div class="t m0 xb9 h25 y144 ff3 fsa fc2 sc2 ls0 ws0">　</div><div class="t m0 x165 h24 y145 ff2 fsa fc2 sc2 ls0 ws0">ｏ<span class="_ _9"></span>ｆ</div><div class="t m0 x45 h25 y144 ff3 fsa fc2 sc2 ls0 ws0">　</div><div class="t m0 x15f h24 y145 ff2 fsa fc2 sc2 ls0 ws0">ｈ</div><div class="t m0 x10a h24 y146 ff2 fsa fc2 sc2 ls0 ws0">ｙ</div><div class="t m0 x61 h24 y145 ff2 fsa fc2 sc2 ls0 ws0">ｂ<span class="_ _7"></span>ｒ<span class="_ _b"></span>ｉ<span class="_ _9"></span>ｄ</div><div class="t m0 x35 h25 y144 ff3 fsa fc2 sc2 ls0 ws0">　</div><div class="t m0 x1a6 h24 y145 ff2 fsa fc2 sc2 ls0 ws0">Ｄ<span class="_ _5"></span>Ｎ<span class="_ _c"></span>Ｎ</div><div class="t m7 x3a h24 y144 ff2 fsa fc2 sc2 ls0 ws0">－</div><div class="t m0 x1a7 h24 y145 ff2 fsa fc2 sc2 ls0 ws0">Ｈ<span class="_ _d"></span>ＭＭ</div><div class="t m0 x12f h24 y147 ff2 fsa fc2 sc2 ls0 ws0">ｂ<span class="_ _8"></span>ａ<span class="_ _7"></span>ｓ<span class="_ _7"></span>ｅ<span class="_ _8"></span>ｄ</div><div class="t m0 x1a8 h25 y148 ff3 fsa fc2 sc2 ls0 ws0">　</div><div class="t m0 x157 h24 y149 ff2 fsa fc2 sc2 ls0 ws0">ｓ</div><div class="t m0 xd6 h24 y14a ff2 fsa fc2 sc2 ls0 ws0">ｐ</div><div class="t m0 x131 h24 y149 ff2 fsa fc2 sc2 ls0 ws0">ｅ<span class="_ _7"></span>ｅ<span class="_ _9"></span>ｃ<span class="_ _1"></span>ｈ</div><div class="t m0 x1a1 h25 y148 ff3 fsa fc2 sc2 ls0 ws0">　</div><div class="t m0 xe4 h24 y149 ff2 fsa fc2 sc2 ls0 ws0">ｒ<span class="_ _9"></span>ｅ<span class="_ _7"></span>ｃ<span class="_ _7"></span>ｏ</div><div class="t m0 x160 h24 y14a ff2 fsa fc2 sc2 ls0 ws0">ｇ</div><div class="t m0 x1a9 h24 y149 ff2 fsa fc2 sc2 ls0 ws0">ｎ<span class="_ _9"></span>ｉ<span class="_ _13"></span>ｔ<span class="_ _b"></span>ｉ<span class="_ _a"></span>ｏ<span class="_ _1"></span>ｎ</div><div class="t m0 x37 h25 y148 ff3 fsa fc2 sc2 ls0 ws0">　</div><div class="t m0 x38 h24 y149 ff2 fsa fc2 sc2 ls0 ws0">ｓ</div><div class="t m0 x43 h24 y14a ff2 fsa fc2 sc2 ls0 ws0">ｙ</div><div class="t m0 x1aa h24 y149 ff2 fsa fc2 sc2 ls0 ws0">ｓ<span class="_ _9"></span>ｔ<span class="_ _a"></span>ｅ<span class="_ _6"></span>ｍ</div><div class="t m0 x57 hd y14b ff8 fs5 fc2 sc2 lsa ws0">最<span class="_ _16"></span>初<span class="_ _d"></span>主<span class="_ _d"></span>流<span class="_ _16"></span>的<span class="_ _d"></span>深层神经网络是最简单的</div><div class="t m0 x113 h11 y14c ff2 fs5 fc2 sc2 ls0 ws0">Ｆ<span class="_ _3"></span>Ｎ<span class="_ _5"></span>Ｎ</div><div class="t m0 x19d h16 y14d ff10 fs5 fc2 sc2 ls0 ws0">。</div><div class="t m0 x51 hd y14b ff8 fs5 fc2 sc2 ls0 ws0">对<span class="_ _17"></span>于</div><div class="t m0 x139 h11 y14c ff2 fs5 fc2 sc2 ls0 ws0">１</div><div class="t m0 x27 hd y14e ff8 fs5 fc2 sc2 ls0 ws0">个<span class="_ _f"></span>包<span class="_ _f"></span>含</div><div class="t m6 x58 h1e y14f ffe fs9 fc2 sc2 ls0 ws0">Ｌ</div><div class="t m0 x19e hd y14e ff8 fs5 fc2 sc2 ls0 ws0">个<span class="_ _f"></span>隐<span class="_ _f"></span>层<span class="_ _e"></span>的</div><div class="t m0 xc5 h11 y14f ff2 fs5 fc2 sc2 ls0 ws0">Ｆ<span class="_ _3"></span>Ｎ<span class="_ _5"></span>Ｎ</div><div class="t m0 x73 h16 y150 ff10 fs5 fc2 sc2 ls0 ws0">，</div><div class="t m0 xc7 hd y14e ff8 fs5 fc2 sc2 ls4 ws0">其整个模型可以表示为</div><div class="t m6 x177 h1e y151 ffe fs9 fc2 sc2 ls0 ws0">ｈ</div><div class="t m0 x19a h1a y152 ff2 fs7 fc2 sc2 ls0 ws0">０</div><div class="t m0 x66 h11 y153 ffe fs5 fc2 sc2 ls0 ws0">＝</div><div class="t m6 x7d h1e y151 ffe fs9 fc2 sc2 ls0 ws0">Ｘ</div><div class="t m0 x1e h16 y154 ff10 fs5 fc2 sc2 ls0 ws0">（</div><div class="t m0 x155 h11 y151 ff2 fs5 fc2 sc2 ls0 ws0">１</div><div class="t m0 x13a h16 y154 ff10 fs5 fc2 sc2 ls0 ws0">）</div><div class="t m6 x111 h1e y155 ffe fs9 fc2 sc2 ls0 ws0">ｈ</div><div class="t m8 xf h26 y156 ffe fsb fc2 sc2 ls0 ws0">ｌ</div><div class="t m0 x163 h11 y157 ffe fs5 fc2 sc2 ls0 ws0">＝</div><div class="t m6 xab h1e y158 ffe fs9 fc2 sc2 ls0 ws0">ｆ</div><div class="t m0 x25 h16 y159 ff10 fs5 fc2 sc2 ls0 ws0">（</div><div class="t m6 xc4 h1e y155 ffe fs9 fc2 sc2 ls0 ws0">Ｗ</div><div class="t m8 xa3 h26 y156 ffe fsb fc2 sc2 ls0 ws0">ｌ</div><div class="t m6 x12 h1e y155 ffe fs9 fc2 sc2 ls0 ws0">ｈ</div><div class="t m8 xad h26 y156 ffe fsb fc2 sc2 ls0 ws0">ｌ</div><div class="t m0 x72 h1a y15a ffe fs7 fc2 sc2 ls0 ws0">－</div><div class="t m0 x182 h1a y156 ff2 fs7 fc2 sc2 ls0 ws0">１</div><div class="t m0 x91 h11 y157 ffe fs5 fc2 sc2 ls0 ws0">＋</div><div class="t m6 xa5 h1e y155 ffe fs9 fc2 sc2 ls0 ws0">ｂ</div><div class="t m8 x89 h26 y156 ffe fsb fc2 sc2 ls0 ws0">ｌ</div><div class="t m0 x135 h16 y159 ff10 fs5 fc2 sc2 ls0 ws0">）</div><div class="t m0 x7e he y15b ff3 fs5 fc2 sc2 ls0 ws0">　<span class="_ _f"></span>　</div><div class="t m0 x175 h11 y155 ff2 fs5 fc2 sc2 ls0 ws0">１</div><div class="t m0 xb1 he y15c ff3 fs5 fc2 sc2 ls0 ws0">≤</div><div class="t m6 x141 h1e y155 ffe fs9 fc2 sc2 ls0 ws0">ｌ</div><div class="t m0 x100 he y15c ff3 fs5 fc2 sc2 ls0 ws0">≤</div><div class="t m6 x6a h1e y155 ffe fs9 fc2 sc2 ls0 ws0">Ｌ</div><div class="t m0 x1e h16 y159 ff10 fs5 fc2 sc2 ls0 ws0">（</div><div class="t m0 x155 h11 y155 ff2 fs5 fc2 sc2 ls0 ws0">２</div><div class="t m0 x13a h16 y159 ff10 fs5 fc2 sc2 ls0 ws0">）</div><div class="t m6 x128 h1e y15d ffe fs9 fc2 sc2 ls0 ws0">ｙ</div><div class="t m0 x2b h11 y15e ffe fs5 fc2 sc2 ls0 ws0">＝</div><div class="t m0 x5c h11 y15f ff2 fs5 fc2 sc2 ls0 ws0">Ｓ<span class="_ _8"></span>ｏ<span class="_ _12"></span>ｆ<span class="_ _15"></span>ｔ<span class="_ _2"></span>ｍ<span class="_ _3"></span>ａ<span class="_ _7"></span>ｘ</div><div class="t m0 x93 h16 y160 ff10 fs5 fc2 sc2 ls0 ws0">（</div><div class="t m6 xfd h1e y15f ffe fs9 fc2 sc2 ls0 ws0">ｗ</div><div class="t m8 x135 h26 y161 ffe fsb fc2 sc2 ls0 ws0">Ｌ</div><div class="t m0 x74 h1a y162 ffe fs7 fc2 sc2 ls0 ws0">＋</div><div class="t m0 x12c h1a y161 ff2 fs7 fc2 sc2 ls0 ws0">１</div><div class="t m0 x122 h27 y163 ff3 fs7 fc2 sc2 ls0 ws0">　</div><div class="t m6 x122 h1e y15f ffe fs9 fc2 sc2 ls0 ws0">ｈ</div><div class="t m8 x17 h26 y161 ffe fsb fc2 sc2 ls0 ws0">Ｌ</div><div class="t m0 x175 h11 y15e ffe fs5 fc2 sc2 ls0 ws0">＋</div><div class="t m6 x124 h1e y15f ffe fs9 fc2 sc2 ls0 ws0">ｂ</div><div class="t m8 x97 h26 y161 ffe fsb fc2 sc2 ls0 ws0">Ｌ</div><div class="t m0 xb2 h1a y162 ffe fs7 fc2 sc2 ls0 ws0">＋</div><div class="t m0 xa9 h1a y161 ff2 fs7 fc2 sc2 ls0 ws0">１</div><div class="t m0 x1ab h16 y160 ff10 fs5 fc2 sc2 ls0 ws0">）<span class="_ _19"> </span>（</div><div class="t m0 x155 h11 y15f ff2 fs5 fc2 sc2 ls0 ws0">３</div><div class="t m0 x13a h16 y160 ff10 fs5 fc2 sc2 ls0 ws0">）</div><div class="t m0 x27 hd y164 ff8 fs5 fc2 sc2 ls0 ws0">式<span class="_ _f"></span>中</div><div class="t m0 x57 h16 y165 ff10 fs5 fc2 sc2 ls0 ws0">：</div><div class="t m6 x11f h1e y166 ffe fs9 fc2 sc2 ls0 ws0">Ｘ</div><div class="t m0 x5 hd y164 ff8 fs5 fc2 sc2 ls6 ws0">表<span class="_ _d"></span>示输<span class="_ _16"></span>入的语音声学特征</div><div class="t m0 x31 h16 y165 ff10 fs5 fc2 sc2 ls0 ws0">；<span class="_ _7"></span>｛</div><div class="t m6 x145 h1e y166 ffe fs9 fc2 sc2 ls0 ws0">Ｗ</div><div class="t m8 x19b h26 y167 ffe fsb fc2 sc2 ls0 ws0">ｌ</div><div class="t m0 x80 h16 y165 ff10 fs5 fc2 sc2 ls0 ws0">，</div><div class="t m6 xa9 h1e y166 ffe fs9 fc2 sc2 ls0 ws0">ｂ</div><div class="t m8 x1ab h26 y167 ffe fsb fc2 sc2 ls0 ws0">ｌ</div><div class="t m0 x6d h16 y165 ff10 fs5 fc2 sc2 ls0 ws0">｝</div><div class="t m0 x183 hd y164 ff8 fs5 fc2 sc2 ls0 ws0">分<span class="_ _e"></span>别<span class="_ _e"></span>表<span class="_ _17"> </span>示</div><div class="t m6 x1ac h1e y166 ffe fs9 fc2 sc2 ls0 ws0">ｌ</div><div class="t m0 x79 hd y164 ff8 fs5 fc2 sc2 ls0 ws0">层</div><div class="t m0 x27 hd y168 ff8 fs5 fc2 sc2 ls0 ws0">的<span class="_ _f"></span>连<span class="_ _f"></span>接<span class="_ _e"></span>权<span class="_ _e"></span>重<span class="_ _17"> </span>和<span class="_ _17"> </span>偏<span class="_ _1a"> </span>量</div><div class="t m0 xc5 h16 y169 ff10 fs5 fc2 sc2 ls0 ws0">；</div><div class="t m6 xac h1e y16a ffe fs9 fc2 sc2 ls0 ws0">ｆ</div><div class="t m0 xee h16 y169 ff10 fs5 fc2 sc2 ls0 ws0">（<span class="_ _1"></span>·<span class="_ _e"> </span>）</div><div class="t m0 x93 hd y16b ff8 fs5 fc2 sc2 lsd ws0">表示隐<span class="_ _d"></span>层的非线性激活函数</div><div class="t m0 x1f h16 y169 ff10 fs5 fc2 sc2 ls0 ws0">。</div><div class="t m0 x27 hd y16c ff8 fs5 fc2 sc2 ls0 ws0">输<span class="_ _f"></span>出<span class="_ _f"></span>层<span class="_ _e"></span>采<span class="_ _f"></span>用</div><div class="t m0 xe h11 y16d ff2 fs5 fc2 sc2 ls0 ws0">Ｓ<span class="_ _8"></span>ｏ<span class="_ _b"></span>ｆ<span class="_ _b"></span>ｔ<span class="_ _1"></span>ｍ<span class="_ _6"></span>ａ<span class="_ _7"></span>ｘ</div><div class="t m0 x18c hd y16c ff8 fs5 fc2 sc2 lsa ws0">函数得到每个建模单元的后验概<span class="_ _10"></span>率</div><div class="t m0 x27 hd y16e ff8 fs5 fc2 sc2 ls0 ws0">输<span class="_ _f"></span>出</div><div class="t m0 x57 h16 y16f ff10 fs5 fc2 sc2 ls0 ws0">。</div><div class="t m0 x58 hd y16e ff8 fs5 fc2 sc2 lsa ws0">通<span class="_ _16"></span>过<span class="_ _d"></span>网<span class="_ _d"></span>络<span class="_ _d"></span>的输出和对应的标注可以设计相应的优</div><div class="t m0 x27 hd y170 ff8 fs5 fc2 sc2 ls0 ws0">化<span class="_ _f"></span>目<span class="_ _17"> </span>标<span class="_ _1b"> </span>函<span class="_ _1b"> </span>数<span class="_ _1c"> </span>进<span class="_ _1c"> </span>行<span class="_ _1b"> </span>模<span class="_ _1b"> </span>型<span class="_ _1b"> </span>的<span class="_ _1c"> </span>优<span class="_ _1b"> </span>化</div><div class="t m0 x7e h16 y171 ff10 fs5 fc2 sc2 ls0 ws0">。</div><div class="t m0 x30 hd y170 ff8 fs5 fc2 sc2 ls0 ws0">交<span class="_ _1b"> </span>叉<span class="_ _1c"> </span>熵</div><div class="t m0 x100 h16 y171 ff10 fs5 fc2 sc2 ls0 ws0">（</div><div class="t m0 x1ad h11 y172 ff2 fs5 fc2 sc2 ls0 ws0">Ｃ<span class="_ _8"></span>ｒ<span class="_ _9"></span>ｏ<span class="_ _9"></span>ｓ<span class="_ _a"></span>ｓ</div><div class="t m4 x150 h11 y173 ff2 fs5 fc2 sc2 ls0 ws0">－</div><div class="t m0 x19d h11 y172 ff2 fs5 fc2 sc2 ls0 ws0">ｅ<span class="_ _7"></span>ｎ<span class="_ _a"></span>ｔ<span class="_ _b"></span>ｒ<span class="_ _9"></span>ｏ</div><div class="t m0 x1d h11 y174 ff2 fs5 fc2 sc2 ls0 ws0">ｐ<span class="_ _7"></span>ｙ</div><div class="t m0 x1f h16 y171 ff10 fs5 fc2 sc2 ls0 ws0">，</div><div class="t m0 x6e h11 y175 ff2 fs5 fc2 sc2 ls0 ws0">Ｃ<span class="_ _2"></span>Ｅ</div><div class="t m0 x84 h16 y176 ff10 fs5 fc2 sc2 ls0 ws0">）</div><div class="t m0 x4 hd y177 ff8 fs5 fc2 sc2 ls6 ws0">准<span class="_ _d"></span>则经<span class="_ _16"></span>常被<span class="_ _d"></span>用作优化目标函数</div><div class="t m0 x8a h16 y176 ff10 fs5 fc2 sc2 ls0 ws0">。</div><div class="t m0 x124 h11 y175 ff2 fs5 fc2 sc2 ls0 ws0">Ｃ<span class="_ _2"></span>Ｅ</div><div class="t m0 x13e hd y177 ff8 fs5 fc2 sc2 ls8 ws0">用来衡量目标输</div><div class="t m0 x27 hd y178 ff8 fs5 fc2 sc2 ls4 ws0">出概率分布和实际输出概<span class="ls5">率分布之间的相似程度</span></div><div class="t m0 x16d h16 y179 ff10 fs5 fc2 sc2 ls0 ws0">，</div><div class="t m0 xde hd y178 ff8 fs5 fc2 sc2 ls0 ws0">其<span class="_ _e"></span>值<span class="_ _e"></span>熵</div><div class="t m0 x27 hd y17a ff8 fs5 fc2 sc2 ls0 ws0">越<span class="_ _f"></span>小<span class="_ _f"></span>相<span class="_ _e"></span>似<span class="_ _f"></span>程<span class="_ _f"></span>度<span class="_ _e"></span>越<span class="_ _17"> </span>高</div><div class="t m0 x7c h16 y17b ff10 fs5 fc2 sc2 ls0 ws0">，</div><div class="t m0 x185 hd y17a ff8 fs5 fc2 sc2 lsa ws0">从而模型的性能也就越好</div><div class="t m0 x83 h16 y17b ff10 fs5 fc2 sc2 ls0 ws0">。</div><div class="t m0 x16d hd y17a ff8 fs5 fc2 sc2 ls0 ws0">基<span class="_ _17"></span>于</div><div class="t m0 x184 h11 y17c ff2 fs5 fc2 sc2 ls0 ws0">Ｃ<span class="_ _2"></span>Ｅ</div><div class="t m0 x27 hd y17d ff8 fs5 fc2 sc2 ls4 ws0">准则的优化目标函数为</div><div class="t m6 x1ae h1e y17e ffe fs9 fc2 sc2 ls0 ws0">Ｆ</div><div class="t m0 x129 h1a y17f ff2 fs7 fc2 sc2 ls0 ws0">Ｃ<span class="_ _c"></span>Ｅ</div><div class="t m0 x2c h16 y180 ff10 fs5 fc2 sc2 ls0 ws0">（</div><div class="t m6 x67 h1e y17e ffe fs9 fc2 sc2 ls0 ws0">Ｗ</div><div class="t m0 x71 h16 y180 ff10 fs5 fc2 sc2 ls0 ws0">）</div><div class="t m0 x18a h11 y181 ffe fs5 fc2 sc2 ls0 ws0">＝<span class="_ _f"></span>－</div><div class="t m0 x93 h28 y182 ff3 fsc fc2 sc2 ls0 ws0">∑</div><div class="t m9 x66 h29 y183 ffe fsd fc2 sc2 ls0 ws0">Ｎ</div><div class="t m9 x93 h29 y184 ffe fsd fc2 sc2 ls0 ws0">ｒ</div><div class="t m0 x66 h2a y185 ffe fse fc2 sc2 ls0 ws0">＝</div><div class="t m0 x17c h2a y184 ff2 fse fc2 sc2 ls0 ws0">１</div><div class="t m0 xc8 h28 y182 ff3 fsc fc2 sc2 ls0 ws0">∑</div><div class="t m9 x12c h29 y186 ffe fsd fc2 sc2 ls0 ws0">Ｔ</div><div class="t ma xf0 h2b y187 ffe fsf fc2 sc2 ls0 ws0">ｒ</div><div class="t m9 x74 h29 y184 ffe fsd fc2 sc2 ls0 ws0">ｔ</div><div class="t m0 x12c h2a y185 ffe fse fc2 sc2 ls0 ws0">＝</div><div class="t m0 x16 h2a y184 ff2 fse fc2 sc2 ls0 ws0">１</div><div class="t m0 xfe h11 y17e ff2 fs5 fc2 sc2 ls0 ws0">ｌ<span class="_ _b"></span>ｏ</div><div class="t m0 x7f h11 y188 ff2 fs5 fc2 sc2 ls0 ws0">ｇ</div><div class="t m6 x1af h1e y189 ffe fs9 fc2 sc2 ls0 ws0">ｙ</div><div class="t m8 x112 h26 y17f ffe fsb fc2 sc2 ls0 ws0">ｒ<span class="_ _4"></span>ｔ</div><div class="t m0 xb2 h16 y180 ff10 fs5 fc2 sc2 ls0 ws0">（</div><div class="t m6 x80 h1e y17e ffe fs9 fc2 sc2 ls0 ws0">ｓ</div><div class="t m8 x13e h26 y17f ffe fsb fc2 sc2 ls0 ws0">ｒ<span class="_ _4"></span>ｔ</div><div class="t m0 x6d h16 y180 ff10 fs5 fc2 sc2 ls0 ws0">）<span class="_ _1d"> </span>（</div><div class="t m0 x155 h11 y17e ff2 fs5 fc2 sc2 ls0 ws0">４</div><div class="t m0 x13a h16 y180 ff10 fs5 fc2 sc2 ls0 ws0">）</div><div class="t m0 x27 hd y18a ff8 fs5 fc2 sc2 ls0 ws0">式<span class="_ _f"></span>中</div><div class="t m0 x57 h16 y18b ff10 fs5 fc2 sc2 ls0 ws0">：</div><div class="t m6 xb h1e y18c ffe fs9 fc2 sc2 ls0 ws0">ｙ</div><div class="t m8 x120 h26 y18d ffe fsb fc2 sc2 ls0 ws0">ｒ<span class="_ _4"></span>ｔ</div><div class="t m0 x5 h16 y18b ff10 fs5 fc2 sc2 ls0 ws0">（</div><div class="t m6 x6 h1e y18e ffe fs9 fc2 sc2 ls0 ws0">ｓ</div><div class="t m0 x110 h16 y18b ff10 fs5 fc2 sc2 ls0 ws0">）</div><div class="t m0 x127 hd y18a ff8 fs5 fc2 sc2 ls0 ws0">表<span class="_ _f"></span>示<span class="_ _f"></span>在</div><div class="t m6 x1b0 h1e y18e ffe fs9 fc2 sc2 ls0 ws0">ｔ</div><div class="t m0 xa hd y18a ff8 fs5 fc2 sc2 ls0 ws0">时<span class="_ _f"></span>刻<span class="_ _f"></span>第</div><div class="t m6 x93 h1e y18e ffe fs9 fc2 sc2 ls0 ws0">ｒ</div><div class="t m0 x12b hd y18a ff8 fs5 fc2 sc2 ls0 ws0">句<span class="_ _f"></span>话<span class="_ _f"></span>在<span class="_ _e"></span>状<span class="_ _f"></span>态</div><div class="t m6 x1a h1e y18e ffe fs9 fc2 sc2 ls0 ws0">ｓ</div><div class="t m0 x195 hd y18a ff8 fs5 fc2 sc2 ls0 ws0">下<span class="_ _f"></span>对<span class="_ _f"></span>应</div><div class="t m0 xf3 h11 y18e ff2 fs5 fc2 sc2 ls0 ws0">Ｓ<span class="_ _8"></span>ｏ<span class="_ _b"></span>ｆ<span class="_ _b"></span>ｔ<span class="_ _1"></span>ｍ<span class="_ _6"></span>ａ<span class="_ _7"></span>ｘ</div><div class="t m0 x12e hd y18a ff8 fs5 fc2 sc2 ls0 ws0">层<span class="_ _f"></span>的<span class="_ _f"></span>输<span class="_ _e"></span>出</div><div class="t m0 xd3 h16 y18b ff10 fs5 fc2 sc2 ls0 ws0">，</div><div class="t m6 xf6 h1e y18e ffe fs9 fc2 sc2 ls0 ws0">ｓ</div><div class="t m8 x14e h26 y18d ffe fsb fc2 sc2 ls0 ws0">ｒ<span class="_ _4"></span>ｔ</div><div class="t m0 x16c hd y18a ff8 fs5 fc2 sc2 ls0 ws0">表<span class="_ _f"></span>示</div><div class="t m6 x1b1 h1e y18e ffe fs9 fc2 sc2 ls0 ws0">Ｘ</div><div class="t m8 x198 h26 y18d ffe fsb fc2 sc2 ls0 ws0">ｒ<span class="_ _4"></span>ｔ</div><div class="t m0 x45 hd y18a ff8 fs5 fc2 sc2 ls0 ws0">对<span class="_ _f"></span>应<span class="_ _f"></span>的<span class="_ _e"></span>标<span class="_ _f"></span>注</div><div class="t m0 x1b2 h16 y18b ff10 fs5 fc2 sc2 ls0 ws0">。</div><div class="t m0 xc3 h11 y18f ff2 fs5 fc2 sc2 ls0 ws0">Ｃ<span class="_ _2"></span>Ｅ</div><div class="t m0 x5 hd y190 ff8 fs5 fc2 sc2 ls4 ws0">是定义在帧级别上的优化准则</div><div class="t m0 x97 h16 y191 ff10 fs5 fc2 sc2 ls0 ws0">，</div><div class="t m0 x1b3 hd y190 ff8 fs5 fc2 sc2 ls0 ws0">由<span class="_ _f"></span>于<span class="_ _f"></span>语<span class="_ _e"></span>音<span class="_ _f"></span>信<span class="_ _f"></span>号<span class="_ _e"></span>是<span class="_ _e"></span>一<span class="_ _e"></span>个<span class="_ _e"></span>时<span class="_ _17"></span>序<span class="_ _e"></span>信<span class="_ _17"></span>号</div><div class="t m0 x119 h16 y191 ff10 fs5 fc2 sc2 ls0 ws0">，</div><div class="t m0 x105 hd y190 ff8 fs5 fc2 sc2 ls6 ws0">所以更为合适的优化准则应该</div><div class="t m0 x27 hd y192 ff8 fs5 fc2 sc2 ls4 ws0">是定义在整个<span class="ls9">序<span class="_ _d"></span>列上的优化准则</span></div><div class="t m0 x17 h16 y193 ff10 fs5 fc2 sc2 ls0 ws0">。</div><div class="t m0 x75 hd y192 ff8 fs5 fc2 sc2 ls0 ws0">文<span class="_ _17"> </span>献</div><div class="t m0 x13e h16 y193 ff10 fs5 fc2 sc2 ls0 ws0">［</div><div class="t m0 x32 h11 y194 ff2 fs5 fc2 sc2 ls0 ws0">９</div><div class="t m0 x176 h16 y193 ff10 fs5 fc2 sc2 ls0 ws0">］</div><div class="t m0 x114 hd y192 ff8 fs5 fc2 sc2 lsd ws0">中对比<span class="_ _d"></span>了不同句子级的区分性准则</div><div class="t m0 x10a h16 y193 ff10 fs5 fc2 sc2 ls0 ws0">，</div><div class="t m0 x33 hd y192 ff8 fs5 fc2 sc2 lsd ws0">包括最大互信息量</div><div class="t m0 x27 h16 y195 ff10 fs5 fc2 sc2 ls0 ws0">（</div><div class="t m0 x50 h11 y196 ff2 fs5 fc2 sc2 ls0 ws0">Ｍ<span class="_ _6"></span>ａ<span class="_ _8"></span>ｘ<span class="_ _b"></span>ｉ<span class="_ _1"></span>ｍ<span class="_ _6"></span>ｕ<span class="_ _6"></span>ｍ</div><div class="t m0 x1b4 he y197 ff3 fs5 fc2 sc2 ls0 ws0">　</div><div class="t m0 x62 h11 y196 ff2 fs5 fc2 sc2 ls0 ws0">ｍ<span class="_ _6"></span>ｕ<span class="_ _9"></span>ｔ<span class="_ _9"></span>ｕ<span class="_ _9"></span>ａ<span class="_ _b"></span>ｌ</div><div class="t m0 x26 he y197 ff3 fs5 fc2 sc2 ls0 ws0">　</div><div class="t m0 x5c h11 y196 ff2 fs5 fc2 sc2 ls0 ws0">ｉ<span class="_ _b"></span>ｎ<span class="_ _a"></span>ｆ<span class="_ _a"></span>ｏ<span class="_ _9"></span>ｒ<span class="_ _4"></span>ｍ<span class="_ _6"></span>ａ<span class="_ _b"></span>ｔ<span class="_ _13"></span>ｉ<span class="_ _b"></span>ｏ<span class="_ _7"></span>ｎ</div><div class="t m0 x74 h16 y195 ff10 fs5 fc2 sc2 ls0 ws0">，</div><div class="t m0 x123 h11 y196 ff2 fs5 fc2 sc2 ls0 ws0">ＭＭ<span class="_ _2"></span>Ｉ</div><div class="t m0 x112 h16 y195 ff10 fs5 fc2 sc2 ls0 ws0">）<span class="_ _8"></span>、</div><div class="t m0 x19 hd y198 ff8 fs5 fc2 sc2 ls4 ws0">最小音素错误率</div><div class="t m0 x53 h16 y195 ff10 fs5 fc2 sc2 ls0 ws0">（</div><div class="t m0 x13b h11 y196 ff2 fs5 fc2 sc2 ls0 ws0">Ｍ<span class="_ _4"></span>ｉ<span class="_ _b"></span>ｎ<span class="_ _b"></span>ｉ<span class="_ _1"></span>ｍ<span class="_ _6"></span>ｕ<span class="_ _6"></span>ｍ</div><div class="t m0 x9c he y197 ff3 fs5 fc2 sc2 ls0 ws0">　</div><div class="t m0 x170 h11 y199 ff2 fs5 fc2 sc2 ls0 ws0">ｐ</div><div class="t m0 x1b5 h11 y19a ff2 fs5 fc2 sc2 ls0 ws0">ｈ<span class="_ _8"></span>ｏ<span class="_ _7"></span>ｎ<span class="_ _7"></span>ｅ</div><div class="t m0 xd5 he y197 ff3 fs5 fc2 sc2 ls0 ws0">　</div><div class="t m0 x193 h11 y196 ff2 fs5 fc2 sc2 ls0 ws0">ｅ<span class="_ _a"></span>ｒ<span class="_ _b"></span>ｒ<span class="_ _9"></span>ｏ<span class="_ _a"></span>ｒ</div><div class="t m0 xf9 h16 y195 ff10 fs5 fc2 sc2 ls0 ws0">，</div><div class="t m0 x109 h11 y196 ff2 fs5 fc2 sc2 ls0 ws0">Ｍ<span class="_ _5"></span>Ｐ<span class="_ _2"></span>Ｅ</div><div class="t m0 xfa h16 y195 ff10 fs5 fc2 sc2 ls0 ws0">）<span class="_ _8"></span>、</div><div class="t m0 x40 hd y198 ff8 fs5 fc2 sc2 ls0 ws0">状<span class="_ _e"></span>态<span class="_ _e"></span>级<span class="_ _e"></span>最<span class="_ _e"></span>小<span class="_ _e"></span>贝</div><div class="t m0 x27 hd y19b ff8 fs5 fc2 sc2 ls0 ws0">叶<span class="_ _e"></span>斯<span class="_ _f"></span>风<span class="_ _e"></span>险</div><div class="t m0 xa1 h16 y19c ff10 fs5 fc2 sc2 ls0 ws0">（</div><div class="t m0 x110 h11 y19d ff2 fs5 fc2 sc2 ls0 ws0">Ｓ<span class="_ _9"></span>ｔ<span class="_ _b"></span>ａ<span class="_ _a"></span>ｔ<span class="_ _b"></span>ｅ</div><div class="t m4 x2a h11 y19e ff2 fs5 fc2 sc2 ls0 ws0">－</div><div class="t m0 x129 h11 y19d ff2 fs5 fc2 sc2 ls0 ws0">ｌ<span class="_ _b"></span>ｅ<span class="_ _9"></span>ｖ<span class="_ _9"></span>ｅ<span class="_ _b"></span>ｌ</div><div class="t m0 xa3 he y19f ff3 fs5 fc2 sc2 ls0 ws0">　</div><div class="t m0 x18c h11 y19d ff2 fs5 fc2 sc2 ls0 ws0">ｍ<span class="_ _1"></span>ｉ<span class="_ _b"></span>ｎ<span class="_ _b"></span>ｉ<span class="_ _1"></span>ｍ<span class="_ _6"></span>ｕ<span class="_ _6"></span>ｍ</div><div class="t m0 xc8 he y19f ff3 fs5 fc2 sc2 ls0 ws0">　</div><div class="t m0 x7e h11 y19d ff2 fs5 fc2 sc2 ls0 ws0">Ｂ<span class="_ _8"></span>ａ</div><div class="t m0 xfe h11 y1a0 ff2 fs5 fc2 sc2 ls0 ws0">ｙ</div><div class="t m0 x95 h11 y19d ff2 fs5 fc2 sc2 ls0 ws0">ｅ<span class="_ _a"></span>ｓ</div><div class="t m0 x1af he y19f ff3 fs5 fc2 sc2 ls0 ws0">　</div><div class="t m0 x124 h11 y19d ff2 fs5 fc2 sc2 ls0 ws0">ｒ<span class="_ _13"></span>ｉ<span class="_ _13"></span>ｓ<span class="_ _9"></span>ｋ</div><div class="t m0 x1ab h16 y19c ff10 fs5 fc2 sc2 ls0 ws0">，</div><div class="t m0 x17d h11 y19d ff2 fs5 fc2 sc2 ls0 ws0">ｓ<span class="_ _6"></span>Ｍ<span class="_ _5"></span>Ｂ<span class="_ _6"></span>Ｒ</div><div class="t m0 xcc h16 y19c ff10 fs5 fc2 sc2 ls0 ws0">）</div><div class="t m0 x102 hd y19b ff8 fs5 fc2 sc2 ls4 ws0">和增强型最大互信息量</div><div class="t m0 x193 h16 y19c ff10 fs5 fc2 sc2 ls0 ws0">（</div><div class="t m0 x1b6 h11 y19d ff2 fs5 fc2 sc2 ls0 ws0">Ｂ<span class="_ _1"></span>ｏ<span class="_ _7"></span>ｏ<span class="_ _9"></span>ｓ<span class="_ _b"></span>ｔ<span class="_ _b"></span>ｅ<span class="_ _7"></span>ｄ</div><div class="t m0 x1b7 he y19f ff3 fs5 fc2 sc2 ls0 ws0">　</div><div class="t m0 x17e h11 y19d ff2 fs5 fc2 sc2 ls0 ws0">ＭＭ<span class="_ _2"></span>Ｉ</div><div class="t m0 xfb h16 y19c ff10 fs5 fc2 sc2 ls0 ws0">，</div><div class="t m0 x36 h11 y19d ff2 fs5 fc2 sc2 ls0 ws0">Ｂ<span class="_ _5"></span>ＭＭ<span class="_ _2"></span>Ｉ</div><div class="t m0 x148 h16 y19c ff10 fs5 fc2 sc2 ls0 ws0">）<span class="_ _8"></span>，</div><div class="t m0 x149 hd y19b ff8 fs5 fc2 sc2 ls0 ws0">用<span class="_ _f"></span>来</div><div class="t m0 x27 hd y1a1 ff8 fs5 fc2 sc2 ls0 ws0">训<span class="_ _f"></span>练</div><div class="t m0 x10f h11 y1a2 ff2 fs5 fc2 sc2 ls0 ws0">Ｄ<span class="_ _3"></span>Ｎ<span class="_ _c"></span>Ｎ</div><div class="t m4 x110 h11 y1a3 ff2 fs5 fc2 sc2 ls0 ws0">－</div><div class="t m0 x1b4 h11 y1a2 ff2 fs5 fc2 sc2 ls0 ws0">Ｈ<span class="_ _d"></span>ＭＭ</div><div class="t m0 x90 hd y1a1 ff8 fs5 fc2 sc2 ls0 ws0">声<span class="_ _f"></span>学<span class="_ _f"></span>模<span class="_ _e"></span>型</div><div class="t m0 xdb h16 y1a4 ff10 fs5 fc2 sc2 ls0 ws0">。</div><div class="t m0 x12b hd y1a1 ff8 fs5 fc2 sc2 ls0 ws0">结<span class="_ _f"></span>果<span class="_ _f"></span>表<span class="_ _e"></span>明</div><div class="t m0 xa8 h16 y1a4 ff10 fs5 fc2 sc2 ls0 ws0">，</div><div class="t m0 x19 hd y1a1 ff8 fs5 fc2 sc2 ls8 ws0">不<span class="_ _d"></span>同句子级区分性准则可以获得相近的性能</div><div class="t m0 x143 h16 y1a4 ff10 fs5 fc2 sc2 ls0 ws0">，</div><div class="t m0 x1a6 hd y1a1 ff8 fs5 fc2 sc2 ls0 ws0">同<span class="_ _e"></span>时<span class="_ _17"> </span>相<span class="_ _17"></span>比<span class="_ _e"></span>于</div><div class="t m0 xc0 h11 y1a2 ff2 fs5 fc2 sc2 ls0 ws0">Ｃ<span class="_ _2"></span>Ｅ</div><div class="t m0 x27 hd y1a5 ff8 fs5 fc2 sc2 ls4 ws0">准则可以获得大概</div><div class="t m0 x5c h11 y1a6 ff2 fs5 fc2 sc2 ls0 ws0">１<span class="_ _7"></span>０<span class="_ _c"></span>％</div><div class="t m0 xae hd y1a5 ff8 fs5 fc2 sc2 ls4 ws0">的相对性能提升</div><div class="t m0 x1b8 h16 y1a7 ff10 fs5 fc2 sc2 ls0 ws0">。</div><div class="t m0 x17d hd y1a5 ff8 fs5 fc2 sc2 ls4 ws0">句子级区分性准<span class="ls5">则通过引入句子级的来自声学模型</span></div><div class="t m0 x49 h16 y1a7 ff10 fs5 fc2 sc2 ls0 ws0">、</div><div class="t m0 x148 hd y1a5 ff8 fs5 fc2 sc2 ls0 ws0">词<span class="_ _e"></span>典<span class="_ _e"></span>和</div><div class="t m0 x27 hd y1a8 ff8 fs5 fc2 sc2 ls4 ws0">语言模型的约束来调整网络参数</div><div class="t m0 x123 h16 y1a9 ff10 fs5 fc2 sc2 ls0 ws0">，</div><div class="t m0 xff hd y1a8 ff8 fs5 fc2 sc2 ls4 ws0">这些约束以网格的形式存储</div><div class="t m0 x117 h16 y1a9 ff10 fs5 fc2 sc2 ls0 ws0">，</div><div class="t m0 x5f hd y1a8 ff8 fs5 fc2 sc2 ls4 ws0">往往需要占用很大的存储空间</div><div class="t m0 x11d h16 y1a9 ff10 fs5 fc2 sc2 ls0 ws0">。</div><div class="t m0 x148 hd y1a8 ff8 fs5 fc2 sc2 ls0 ws0">针<span class="_ _e"></span>对<span class="_ _e"></span>于</div><div class="t m0 x27 hd y1aa ff8 fs5 fc2 sc2 ls0 ws0">此</div><div class="t m0 x56 h16 y1ab ff10 fs5 fc2 sc2 ls0 ws0">，</div><div class="t m0 x3 hd y1aa ff8 fs5 fc2 sc2 ls0 ws0">文<span class="_ _f"></span>献</div><div class="t m0 x5 h16 y1ab ff10 fs5 fc2 sc2 ls0 ws0">［</div><div class="t m0 x85 h11 y1ac ff2 fs5 fc2 sc2 ls0 ws0">１<span class="_ _7"></span>０</div><div class="t m0 x171 h16 y1ab ff10 fs5 fc2 sc2 ls0 ws0">］</div><div class="t m0 x7 hd y1aa ff8 fs5 fc2 sc2 ls4 ws0">提出了一种与网格无关的</div><div class="t m0 x186 h11 y1ac ff2 fs5 fc2 sc2 ls0 ws0">ＭＭ<span class="_ _2"></span>Ｉ</div><div class="t m0 x77 hd y1aa ff8 fs5 fc2 sc2 ls0 ws0">准<span class="_ _f"></span>则<span class="_ _f"></span>用<span class="_ _e"></span>于</div><div class="t m0 x1e h11 y1ac ff2 fs5 fc2 sc2 ls0 ws0">Ｄ<span class="_ _3"></span>Ｎ<span class="_ _c"></span>Ｎ</div><div class="t m0 x16e hd y1aa ff8 fs5 fc2 sc2 ls4 ws0">的句子级区分性训练</div><div class="t m0 x109 h16 y1ab ff10 fs5 fc2 sc2 ls0 ws0">。</div><div class="t m0 x159 h20 y1ad ffc fs8 fc2 sc2 ls0 ws0">２<span class="_ _5"></span>．<span class="_ _14"></span>２</div><div class="t m0 x3 h21 y1ae ff3 fs8 fc2 sc2 ls0 ws0">　</div><div class="t m0 x6f h22 y1af ff6 fs8 fc2 sc2 lse ws0">基于深度学习的语音识别声学模型的结构或类型</div><div class="t m0 x4 h11 y1b0 ff2 fs5 fc2 sc2 ls0 ws0">２<span class="_ _7"></span>０<span class="_ _8"></span>０<span class="_ _8"></span>９</div><div class="t m0 x4d hd y1b1 ff8 fs5 fc2 sc2 ls0 ws0">年</div><div class="t m0 x133 h16 y1b2 ff10 fs5 fc2 sc2 ls0 ws0">，</div><div class="t m0 x24 h11 y1b0 ff2 fs5 fc2 sc2 ls0 ws0">Ｄ<span class="_ _3"></span>Ｎ<span class="_ _c"></span>Ｎ</div><div class="t m0 x7c hd y1b1 ff8 fs5 fc2 sc2 ls4 ws0">首次被应用到语音识别中</div><div class="t m0 x77 h19 y1b3 ff10 fs7 fc2 sc2 ls0 ws0">［</div><div class="t m0 x1b h1a y1b4 ff2 fs7 fc2 sc2 ls0 ws0">５</div><div class="t m0 x1c h19 y1b3 ff10 fs7 fc2 sc2 ls0 ws0">］</div><div class="t m0 x83 h16 y1b2 ff10 fs5 fc2 sc2 ls0 ws0">，</div><div class="t m0 xf2 hd y1b1 ff8 fs5 fc2 sc2 ls4 ws0">其采用的模型为如图</div><div class="t m0 xf6 h11 y1b0 ff2 fs5 fc2 sc2 ls0 ws0">１</div><div class="t m0 x1b9 hd y1b1 ff8 fs5 fc2 sc2 ls0 ws0">所<span class="_ _f"></span>示<span class="_ _f"></span>的</div><div class="t m0 x11a h11 y1b0 ff2 fs5 fc2 sc2 ls0 ws0">Ｄ<span class="_ _3"></span>Ｎ<span class="_ _c"></span>Ｎ</div><div class="t m4 x1a4 h11 y1b5 ff2 fs5 fc2 sc2 ls0 ws0">－</div><div class="t m0 x174 h11 y1b0 ff2 fs5 fc2 sc2 ls0 ws0">Ｈ<span class="_ _d"></span>ＭＭ</div><div class="t m0 x43 h16 y1b2 ff10 fs5 fc2 sc2 ls0 ws0">。</div><div class="t m0 x11d hd y1b1 ff8 fs5 fc2 sc2 ls0 ws0">当<span class="_ _e"></span>时<span class="_ _e"></span>的<span class="_ _e"></span>实</div><div class="t m0 x1ba h1d y113 ff2 fs8 fc2 sc2 ls0 ws0">３<span class="_ _18"></span>２<span class="_ _18"></span>２</div><div class="t m0 x27 he y116 ff3 fs5 fc2 sc2 ls0 ws0">　</div><div class="t m0 x56 hd y114 ffa fs5 fc2 sc2 ls0 ws0">戴<span class="_ _f"></span>礼<span class="_ _f"></span>荣<span class="_ _1e"> </span>等</div><div class="t m0 xf h2c y119 ff11 fs5 fc2 sc2 ls0 ws0">：</div><div class="t m0 x1ae hd y114 ffa fs5 fc2 sc2 ls4 ws0">基于深度学习的语音识别技术现状与展望</div></div><div class="pi" data-data='{"ctm":[1.830664,0.000000,0.000000,1.830664,0.000000,0.000000]}'></div></div>
java
