﻿---
title: "训练大模型为什么不用来回评审：梯度下降找到的是什么解"
description: "人解决问题要出方案、评审、再改，来回好多轮，新方案也未必最好；大模型训练却好像只管顺着梯度往下走，就走到了一个不错的解。顺着这个疑问往下追：训练里谁在评审，为什么每一步都不保证变好却还是越来越好，为什么不容易卡在半路，继续训能不能让每个参数都最优，以及人的那些来回评审最后去了哪里。"
pubDate: 2026-10-29
tags: ["AI", "梯度下降", "神经网络", "入门"]
hero: "./hero.svg"
clc: "TP183"
refs:
  - id: gpt3
    title: "Language Models are Few-Shot Learners"
    url: https://arxiv.org/abs/2005.14165
    author: Tom B. Brown 等
    site: NeurIPS 2020
    note: "表 2.1：175B 模型批大小 320 万词元，共训练 3000 亿词元，两者相除约 9.4 万步。"
  - id: dauphin
    title: "Identifying and attacking the saddle point problem in high-dimensional non-convex optimization"
    url: https://arxiv.org/abs/1406.2572
    author: Yann Dauphin, Razvan Pascanu, Caglar Gulcehre, Kyunghyun Cho, Surya Ganguli, Yoshua Bengio
    site: NeurIPS 2014
    note: "论点：高维优化的主要障碍不是误差很高的局部极小，而是大量鞍点，它们周围的平地会让训练变慢，看起来像卡住。"
  - id: choromanska
    title: "The Loss Surfaces of Multilayer Networks"
    url: https://arxiv.org/abs/1412.0233
    author: Anna Choromanska, Mikael Henaff, Michael Mathieu, Gérard Ben Arous, Yann LeCun
    site: AISTATS 2015
    note: "在简化假设下论证：网络越大，局部极小的损失越集中在一个很窄的低值带里，差的局部极小越来越少。"
  - id: garipov
    title: "Loss Surfaces, Mode Connectivity, and Fast Ensembling of DNNs"
    url: https://arxiv.org/abs/1802.10026
    author: Timur Garipov, Pavel Izmailov, Dmitrii Podoprikhin, Dmitry Vetrov, Andrew Gordon Wilson
    site: NeurIPS 2018
  - id: draxler
    title: "Essentially No Barriers in Neural Network Energy Landscape"
    url: https://arxiv.org/abs/1803.00885
    author: Felix Draxler, Kambis Veschgini, Manfred Salmhofer, Fred A. Hamprecht
    site: ICML 2018
  - id: zhang
    title: "Understanding deep learning requires rethinking generalization"
    url: https://arxiv.org/abs/1611.03530
    author: Chiyuan Zhang, Samy Bengio, Moritz Hardt, Benjamin Recht, Oriol Vinyals
    site: ICLR 2017
    note: "常见的图像网络可以把随机打乱的标签也背到训练误差为 0。"
  - id: chinchilla
    title: "Training Compute-Optimal Large Language Models"
    url: https://arxiv.org/abs/2203.15556
    author: Jordan Hoffmann 等
    site: NeurIPS 2022
  - id: llama3
    title: "Introducing Meta Llama 3"
    url: https://ai.meta.com/blog/meta-llama-3/
    site: Meta AI
    date: 2024-04-18
    note: "原文：8B 模型按 Chinchilla 算法约 2000 亿词元就够；8B 和 70B 训到 15 万亿词元后仍在按对数线性提升。"
  - id: instructgpt
    title: "Training language models to follow instructions with human feedback"
    url: https://arxiv.org/abs/2203.02155
    author: Long Ouyang 等
    site: NeurIPS 2022
  - id: gao
    title: "Scaling Laws for Reward Model Overoptimization"
    url: https://arxiv.org/abs/2210.10760
    author: Leo Gao, John Schulman, Jacob Hilton
    site: ICML 2023
    note: "对着奖励模型优化时，奖励模型给的分一直涨，用更可靠的评分去看，真实质量先升后降。"
  - id: notes
    title: "本文的实验脚本与日志"
    note: "项目共享目录 `gradient-why-works-research/`：`experiment/mlp.py` 是一隐层网络和手写的反向传播，`e_sgd.py` 小批量锯齿，`e_width.py` 宽度与卡住，`e_stuck.py`、`e_escape_hess.py`、`e_one_at_a_time.py` 卡住的点和加一个神经元，`e_saddle.py` 随机弯曲的比例，`e_overfit.py` 过拟合，`build_figs.py` 生成文中的图和表。只用 numpy。"
---

我们平时解决一个问题，大致是这样走的：先出一个方案，落地看看结果，开会评审符不符合预期，提几条修改意见，改出第二版。第二版也不一定比第一版好，于是再评审、再改。一个方案来来回回倒腾好多轮，最后定下来的也只是“目前最满意的”，谁也不敢说它最优。

训练一个大模型，看起来却完全是另一回事。我们在[《Transformer 是怎么训练的》](/blog/transformer-training-by-hand/)里手算过：算出损失对每个参数的坡度，每个参数朝下坡挪一小步，重复。屏幕上的损失从 0.615 一路降到 0.030，中间没有人开会，也没有人拍板说“这一版不行，推倒重来”。几千亿个参数，靠这么一条规则，就被调到了一个相当好的地方。

这就让人想问：人要反复权衡、反复评审才能做好的事，训练里是怎么省掉的？按梯度走一步，凭什么就是在往对的方向改？走到后来，是不是每个参数都已经调到了最优？如果是，这里面到底用了什么魔法。

本文就顺着这几个疑问，一层一层往下讲。

## 一、训练里的评审

### 现象：看起来没人评审

先把训练的一步，和人改方案的一轮，放在一起看（图 1）。

<figure class="dd dd-figure">
<div class="dd-scroll">
<svg class="dd-svg" viewBox="0 0 800 424" xmlns="http://www.w3.org/2000/svg" role="img" aria-labelledby="gw1-title gw1-desc">
<title id="gw1-title">人做方案的评审循环和训练的评审循环</title>
<desc id="gw1-desc">左边是人：出方案、看结果、开会评审、给出几条修改意见，再改。右边是训练的一步：用当前参数答一批题、算出损失、反向传播得到每个参数的梯度、更新参数，再来一轮。两边都是循环，差别在评审给回来的东西：人给几条意见，训练给出每个参数各自该怎么调。</desc>
<defs>
<marker id="gw1-arrow" markerWidth="8" markerHeight="8" refX="7" refY="4" orient="auto"><path class="dd-head" d="M1 1 L7 4 L1 7"/></marker>
<marker id="gw1-arrow-acc" markerWidth="8" markerHeight="8" refX="7" refY="4" orient="auto"><path class="dd-head-acc" d="M1 1 L7 4 L1 7"/></marker>
</defs>
<rect class="dd-frame" width="800" height="424" rx="10"/>
<rect class="dd-zone" x="16" y="16" width="376" height="392" rx="8"/>
<text class="t-zone" x="32" y="44">人做方案</text>
<rect class="dd-zone" x="408" y="16" width="376" height="392" rx="8"/>
<text class="t-zone" x="424" y="44">训练一步</text>
<line class="dd-edge" x1="182" y1="116" x2="182" y2="134" marker-end="url(#gw1-arrow)"/>
<rect class="dd-card" x="32" y="60" width="300" height="56" rx="6"/>
<text class="t-name" x="46" y="82">出一个方案</text>
<text class="t-note" x="46" y="102">按经验和讨论拍板</text>
<line class="dd-edge" x1="182" y1="192" x2="182" y2="210" marker-end="url(#gw1-arrow)"/>
<rect class="dd-card" x="32" y="136" width="300" height="56" rx="6"/>
<text class="t-name" x="46" y="158">落地看结果</text>
<text class="t-note" x="46" y="178">跑一段时间，收集反馈</text>
<line class="dd-edge" x1="182" y1="268" x2="182" y2="286" marker-end="url(#gw1-arrow)"/>
<rect class="dd-card" x="32" y="212" width="300" height="56" rx="6"/>
<text class="t-name" x="46" y="234">开会评审</text>
<text class="t-note" x="46" y="254">符不符合预期？好还是不好？</text>
<rect class="dd-card" x="32" y="288" width="300" height="56" rx="6"/>
<text class="t-name" x="46" y="310">几条修改意见</text>
<text class="t-note" x="46" y="330">“这里再想想”，改哪、改多少靠人定</text>
<path class="dd-edge" style="fill:none" d="M332 316 H352 Q360 316 360 308 V96 Q360 88 352 88 H334" marker-end="url(#gw1-arrow)"/>
<line class="dd-edge" x1="574" y1="116" x2="574" y2="134" marker-end="url(#gw1-arrow)"/>
<rect class="dd-card" x="424" y="60" width="300" height="56" rx="6"/>
<text class="t-name" x="438" y="82">当前的参数</text>
<text class="t-note" x="438" y="102">GPT-3：1750 亿个数</text>
<line class="dd-edge" x1="574" y1="192" x2="574" y2="210" marker-end="url(#gw1-arrow)"/>
<rect class="dd-card" x="424" y="136" width="300" height="56" rx="6"/>
<text class="t-name" x="438" y="158">答一批题（前向）</text>
<text class="t-note" x="438" y="178">约 320 万个词元，每个都猜下一个</text>
<line class="dd-edge" x1="574" y1="268" x2="574" y2="286" marker-end="url(#gw1-arrow)"/>
<rect class="dd-card" x="424" y="212" width="300" height="56" rx="6"/>
<text class="t-name" x="438" y="234">算损失</text>
<text class="t-note" x="438" y="254">一个数：这批题平均错多少</text>
<rect class="dd-card" x="424" y="288" width="300" height="56" rx="6"/>
<text class="t-name t-acc" x="438" y="310">梯度（反向传播）</text>
<text class="t-note t-acc" x="438" y="330">1750 亿个数：每个参数往哪调、调多少</text>
<path class="dd-edge-acc" style="fill:none" d="M724 316 H744 Q752 316 752 308 V96 Q752 88 744 88 H726" marker-end="url(#gw1-arrow-acc)"/>
<text class="t-say" x="48" y="384">改出来的新方案也不一定更好，</text>
<text class="t-say" x="48" y="404">再评审一轮。一轮几天到几周。</text>
<text class="t-say" x="440" y="384">每个参数减去“学习率 × 梯度”，再来一轮。</text>
<text class="t-say" x="440" y="404">GPT-3 一共转了约 9.4 万轮。</text>
</svg>
</div>
<figcaption>图 1 · 训练也是一轮一轮评审出来的，只是评审意见具体到了每一个参数</figcaption>
</figure>

左边是人。方案落地之后，评审会回答的是“好还是不好”，最多再给几条方向性的意见：“这里再想想”“那里改激进一点”。具体改哪、改多少，还得靠人自己琢磨，所以改完常常发现改歪了。

右边是训练的一步。**损失就是评审。** 它是一个提前写死的分数，比如语言模型的损失就是“下一个词猜错了多少”，用当前参数答一批题，算出这个分数，评审就完成了。然后反向传播把这个分数往回拆，拆出每一个参数的偏导数，也就是梯度：这个参数往大调，损失会涨还是降，涨降有多快。这相当于评审会给 1750 亿个参数各写了一条意见，每条都精确到“往哪边、改多少”。

所以训练不是没有评审，而是**每一步都是一轮评审**。GPT-3 一步要答约 320 万个词元的题，一共走了约 9.4 万步[^gpt3]，也就是被评审了 9.4 万轮。

### 评审意见为什么能这么具体

人的评审给不出这么具体的意见，是因为没人知道“方案”和“结果”之间的公式：改一下预算，口碑会变好多少？只能改了再看。训练里，从参数到损失的每一步运算都是我们自己写的公式，能求导，所以评审可以顺带算出“每个参数对结果的影响”。[上一篇](/blog/transformer-training-by-hand/)第十一节专门比过这一点：只看结果好坏的搜索，参数一多就走不动；能拿到每个参数坡度的梯度下降，步数几乎不随参数个数上涨。

不过，评审意见具体，不代表照着改就一定变好。这正是人担心的那件事：新方案可能还不如旧方案。训练里会不会也这样？

### 现象：每一步并不保证变好

会，而且非常频繁。

我们训练一个小网络来做一个分类题：平面上有两条交错的螺旋，200 个点，每个点属于其中一条，网络要学会判断任意一个点属于哪条。

这个网络由**神经元**搭成，后文的实验都用它。一个神经元只做一件很小的事：把收到的几个数各乘一个系数，加起来，再加一个常数，最后把结果压进 −1 到 1 之间（图 2 左）。乘的系数叫**权重**，表示每个输入有多重要；加的常数叫**偏置**，用来整体抬高或压低；压的那一步用的函数叫 tanh，结果越大越贴近 1，越小越贴近 −1，在 0 附近差不多原样输出。“神经元”这个名字来自它粗略地模仿了脑细胞：收到好几路信号，各有轻重，总量够了才兴奋起来。如果没有最后压的那一步，不管叠多少层，整个网络都只是一次加权求和，只能用一条直线去分两类点；有了它，网络才能画出弯曲的分界线。

把很多个神经元并排放，就是一个网络（图 2 右）。每个神经元都读同一个点的两个坐标，但各有各的权重，于是各自对平面上不同的区域敏感；最后把它们的输出再各乘一个权重（叫这个神经元的**输出权重**）加起来，再加一个偏置，得到一个分数，大于 0 判为第一条螺旋，小于 0 判为另一条。训练要调的参数，就是所有这些权重和偏置。

<figure class="dd dd-figure">
<div class="dd-scroll">
<svg class="dd-svg" viewBox="0 0 800 376" xmlns="http://www.w3.org/2000/svg" role="img" aria-labelledby="gw0-title gw0-desc">
<title id="gw0-title">一个神经元怎么算，以及一排神经元怎么组成网络</title>
<desc id="gw0-desc">左边是一个神经元的计算：输入点 (1, −1)，输入权重 0.8 和 −0.5，偏置 0.1，加权求和得 1.4，再经过 tanh 压成 0.885，下面画着 tanh 曲线。右边是网络：两个坐标连到 64 个神经元，每个神经元的输出再汇到一个分数，大于 0 判为第一条螺旋。每个神经元有 2 个输入权重、1 个偏置、1 个输出权重，总共 257 个参数。</desc>
<defs>
<marker id="gw0-arrow" markerWidth="8" markerHeight="8" refX="7" refY="4" orient="auto"><path class="dd-head" d="M1 1 L7 4 L1 7"/></marker>
<marker id="gw0-arrow-acc" markerWidth="8" markerHeight="8" refX="7" refY="4" orient="auto"><path class="dd-head-acc" d="M1 1 L7 4 L1 7"/></marker>
</defs>
<rect class="dd-frame" width="800" height="376" rx="10"/>
<rect class="dd-zone" x="16" y="16" width="376" height="344" rx="8"/>
<text class="t-zone" x="32" y="44">一个神经元</text>
<rect class="dd-card" x="32" y="64" width="344" height="162" rx="6"/>
<path class="dd-card-head" d="M32.6 96 V70 Q32.6 64.6 38 64.6 H370 Q375.4 64.6 375.4 70 V96 Z"/>
<text class="t-name" x="48" y="85">神经元 1</text><text class="t-note" x="360" y="85" text-anchor="end">读一个点</text>
<text class="t-mono" x="48" y="113">输入 (x₁, x₂)</text><text class="t-val" x="360" y="113" text-anchor="end">(1, −1)</text>
<line class="dd-hair" x1="33" y1="122" x2="375" y2="122"/>
<text class="t-mono t-acc" x="48" y="139">输入权重</text><text class="t-val t-acc" x="360" y="139" text-anchor="end">0.8, −0.5</text>
<line class="dd-hair" x1="33" y1="148" x2="375" y2="148"/>
<text class="t-mono t-acc" x="48" y="165">偏置</text><text class="t-val t-acc" x="360" y="165" text-anchor="end">0.1</text>
<line class="dd-hair" x1="33" y1="174" x2="375" y2="174"/>
<text class="t-mono" x="48" y="191">加权求和</text><text class="t-val" x="360" y="191" text-anchor="end">1×0.8 + (−1)×(−0.5) + 0.1 = 1.4</text>
<line class="dd-hair" x1="33" y1="200" x2="375" y2="200"/>
<text class="t-mono" x="48" y="217">压进 −1 到 1</text><text class="t-val" x="360" y="217" text-anchor="end">tanh(1.4) = 0.885</text>
<line class="dd-hair" x1="72" y1="300.0" x2="352" y2="300.0"/>
<line class="dd-hair" x1="212.0" y1="264" x2="212.0" y2="336"/>
<path class="dd-edge" style="stroke-width:2;fill:none" d="M72.0 336.0 L75.5 336.0 L79.0 336.0 L82.5 336.0 L86.0 335.9 L89.5 335.9 L93.0 335.9 L96.5 335.9 L100.0 335.9 L103.5 335.9 L107.0 335.8 L110.5 335.8 L114.0 335.7 L117.5 335.7 L121.0 335.6 L124.5 335.5 L128.0 335.4 L131.5 335.3 L135.0 335.1 L138.5 334.9 L142.0 334.7 L145.5 334.4 L149.0 334.1 L152.5 333.7 L156.0 333.2 L159.5 332.6 L163.0 331.9 L166.5 331.0 L170.0 330.0 L173.5 328.8 L177.0 327.4 L180.5 325.8 L184.0 323.9 L187.5 321.8 L191.0 319.3 L194.5 316.6 L198.0 313.7 L201.5 310.5 L205.0 307.1 L208.5 303.6 L212.0 300.0 L215.5 296.4 L219.0 292.9 L222.5 289.5 L226.0 286.3 L229.5 283.4 L233.0 280.7 L236.5 278.2 L240.0 276.1 L243.5 274.2 L247.0 272.6 L250.5 271.2 L254.0 270.0 L257.5 269.0 L261.0 268.1 L264.5 267.4 L268.0 266.8 L271.5 266.3 L275.0 265.9 L278.5 265.6 L282.0 265.3 L285.5 265.1 L289.0 264.9 L292.5 264.7 L296.0 264.6 L299.5 264.5 L303.0 264.4 L306.5 264.3 L310.0 264.3 L313.5 264.2 L317.0 264.2 L320.5 264.1 L324.0 264.1 L327.5 264.1 L331.0 264.1 L334.5 264.1 L338.0 264.1 L341.5 264.0 L345.0 264.0 L348.5 264.0 L352.0 264.0"/>
<circle class="dd-fill-acc" cx="261.0" cy="268.1" r="5"/>
<text class="t-note" x="64" y="268.0" text-anchor="end">1</text><text class="t-note" x="64" y="340.0" text-anchor="end">−1</text>
<text class="t-note" x="271.0" y="284.1">1.4 → 0.885</text>
<text class="t-say" x="72.0" y="252">tanh：越大越贴近 1，越小越贴近 −1</text>
<rect class="dd-zone" x="408" y="16" width="376" height="344" rx="8"/>
<text class="t-zone" x="424" y="44">64 个神经元并排</text>
<line class="dd-edge-acc" x1="480" y1="150" x2="582" y2="84"/>
<line class="dd-edge" x1="480" y1="150" x2="582" y2="140"/>
<line class="dd-edge" x1="480" y1="150" x2="582" y2="248"/>
<line class="dd-edge-acc" x1="480" y1="230" x2="582" y2="84"/>
<line class="dd-edge" x1="480" y1="230" x2="582" y2="140"/>
<line class="dd-edge" x1="480" y1="230" x2="582" y2="248"/>
<line class="dd-edge-acc" x1="618" y1="84" x2="680" y2="176"/>
<line class="dd-edge" x1="618" y1="140" x2="680" y2="176"/>
<line class="dd-edge" x1="618" y1="248" x2="680" y2="176"/>
<rect class="dd-card" x="424" y="132" width="56" height="36" rx="6"/><text class="t-mono" x="452" y="155" text-anchor="middle">x₁</text>
<rect class="dd-card" x="424" y="212" width="56" height="36" rx="6"/><text class="t-mono" x="452" y="235" text-anchor="middle">x₂</text>
<circle class="dd-card" cx="600" cy="84" r="18"/><text class="t-mono t-acc" x="600" y="88" text-anchor="middle">1</text>
<circle class="dd-card" cx="600" cy="140" r="18"/><text class="t-mono" x="600" y="144" text-anchor="middle">2</text>
<circle class="dd-card" cx="600" cy="248" r="18"/><text class="t-mono" x="600" y="252" text-anchor="middle">64</text>
<text class="t-name" x="600" y="200" text-anchor="middle">⋮</text>
<rect class="dd-card" x="680" y="158" width="88" height="36" rx="6"/><text class="t-name" x="724" y="181" text-anchor="middle">分数</text>
<text class="t-note" x="724" y="214" text-anchor="middle">大于 0：</text><text class="t-note" x="724" y="232" text-anchor="middle">第一条螺旋</text>
<text class="t-say t-acc" x="424" y="304">每个神经元 4 个参数：2 个输入权重、1 个偏置、</text>
<text class="t-say t-acc" x="424" y="324">1 个输出权重。64 × 4 + 最后 1 个偏置 = 257。</text>
</svg>
</div>
<figcaption>图 2 · 神经元只是“加权求和再压一下”，网络的参数就是这些权重和偏置</figcaption>
</figure>

这里用 64 个神经元，一共 257 个参数。真实的大模型训练不会每步都用全部数据，而是每步随机抽一小批，这里每步只抽 20 个点来算梯度。

每走一步，我们都用全部 200 个点量一次损失（图 3）。本文的实验都只用 numpy 写成，脚本和日志见参考资料[^notes]。

<figure class="dd dd-figure">
<div class="dd-scroll">
<svg class="dd-svg" viewBox="0 0 800 352" xmlns="http://www.w3.org/2000/svg" role="img" aria-labelledby="gw2-title gw2-desc">
<title id="gw2-title">小批量训练时损失的锯齿</title>
<desc id="gw2-desc">一个 64 个神经元的小网络学螺旋分类，每步只看 20 个点。细线是每一步之后全部 200 个点的平均损失，上下乱跳，前 1000 步里有 494 步比上一步更差；粗线是前后 50 步的平均，从 0.70 稳定降到接近 0。</desc>
<defs>
<marker id="gw2-arrow" markerWidth="8" markerHeight="8" refX="7" refY="4" orient="auto"><path class="dd-head" d="M1 1 L7 4 L1 7"/></marker>
<marker id="gw2-arrow-acc" markerWidth="8" markerHeight="8" refX="7" refY="4" orient="auto"><path class="dd-head-acc" d="M1 1 L7 4 L1 7"/></marker>
</defs>
<rect class="dd-frame" width="800" height="352" rx="10"/>
<line class="dd-edge" x1="72" y1="300" x2="768" y2="300"/>
<line class="dd-edge" x1="72" y1="40" x2="72" y2="300"/>
<text class="t-note" x="62" y="304.0" text-anchor="end">0</text>
<text class="t-note" x="62" y="217.3" text-anchor="end">0.4</text>
<line class="dd-hair" x1="72" y1="213.3" x2="768" y2="213.3"/>
<text class="t-note" x="62" y="130.7" text-anchor="end">0.8</text>
<line class="dd-hair" x1="72" y1="126.7" x2="768" y2="126.7"/>
<text class="t-note" x="62" y="44.0" text-anchor="end">1.2</text>
<line class="dd-hair" x1="72" y1="40.0" x2="768" y2="40.0"/>
<text class="t-note" x="72.0" y="320" text-anchor="middle">0</text>
<text class="t-note" x="246.0" y="320" text-anchor="middle">500</text>
<text class="t-note" x="420.0" y="320" text-anchor="middle">1000</text>
<text class="t-note" x="594.0" y="320" text-anchor="middle">1500</text>
<text class="t-note" x="768.0" y="320" text-anchor="middle">2000</text>
<text class="t-note" x="768" y="340" text-anchor="end">第几步</text>
<text class="t-note" x="72" y="26">全部 200 个点的平均损失</text>
<path class="dd-edge" style="stroke-width:0.8;fill:none" d="M72.0 149.2 L72.3 131.6 L72.7 118.7 L73.0 140.2 L73.4 97.6 L73.7 152.5 L74.1 152.0 L74.4 144.8 L74.8 134.1 L75.1 124.1 L75.5 155.1 L75.8 53.7 L76.2 105.5 L76.5 156.9 L76.9 130.0 L77.2 135.6 L77.6 127.0 L77.9 152.3 L78.3 160.8 L78.6 122.1 L79.0 162.7 L79.3 128.4 L79.7 140.3 L80.0 131.3 L80.4 136.3 L80.7 121.2 L81.0 161.5 L81.4 160.3 L81.7 152.4 L82.1 73.3 L82.4 152.7 L82.8 136.0 L83.1 87.7 L83.5 159.7 L83.8 168.7 L84.2 147.8 L84.5 149.6 L84.9 173.1 L85.2 173.6 L85.6 167.0 L85.9 174.9 L86.3 166.4 L86.6 169.5 L87.0 174.5 L87.3 176.1 L87.7 173.9 L88.0 158.7 L88.4 89.4 L88.7 156.4 L89.1 180.8 L89.4 172.9 L89.7 182.8 L90.1 143.2 L90.4 170.1 L90.8 180.9 L91.1 180.2 L91.5 135.8 L91.8 158.8 L92.2 186.8 L92.5 183.3 L92.9 182.0 L93.2 184.4 L93.6 188.3 L93.9 183.4 L94.3 172.6 L94.6 183.1 L95.0 162.5 L95.3 185.8 L95.7 180.3 L96.0 177.1 L96.4 109.2 L96.7 178.5 L97.1 178.6 L97.4 170.7 L97.8 163.3 L98.1 135.3 L98.4 139.6 L98.8 173.5 L99.1 191.6 L99.5 192.3 L99.8 185.8 L100.2 192.7 L100.5 191.7 L100.9 179.7 L101.2 183.8 L101.6 196.3 L101.9 172.6 L102.3 196.4 L102.6 184.6 L103.0 169.1 L103.3 180.2 L103.7 183.4 L104.0 165.1 L104.4 198.2 L104.7 180.5 L105.1 199.6 L105.4 192.2 L105.8 154.9 L106.1 186.9 L106.5 130.1 L106.8 181.9 L107.1 181.9 L107.5 203.4 L107.8 125.7 L108.2 154.9 L108.5 189.3 L108.9 178.9 L109.2 206.8 L109.6 203.6 L109.9 189.9 L110.3 205.9 L110.6 206.5 L111.0 193.1 L111.3 183.6 L111.7 184.2 L112.0 182.8 L112.4 189.8 L112.7 209.3 L113.1 208.5 L113.4 183.7 L113.8 170.7 L114.1 167.6 L114.5 192.8 L114.8 206.7 L115.2 200.3 L115.5 187.1 L115.8 202.6 L116.2 184.0 L116.5 146.1 L116.9 183.2 L117.2 197.6 L117.6 190.0 L117.9 200.7 L118.3 213.6 L118.6 205.2 L119.0 159.3 L119.3 192.6 L119.7 200.1 L120.0 212.6 L120.4 200.2 L120.7 207.3 L121.1 217.1 L121.4 210.2 L121.8 209.1 L122.1 195.2 L122.5 210.5 L122.8 198.0 L123.2 154.4 L123.5 170.7 L123.9 208.7 L124.2 161.6 L124.5 185.3 L124.9 178.6 L125.2 195.7 L125.6 211.6 L125.9 211.1 L126.3 211.1 L126.6 210.1 L127.0 202.3 L127.3 215.5 L127.7 215.3 L128.0 206.9 L128.4 205.6 L128.7 217.7 L129.1 163.1 L129.4 194.8 L129.8 147.9 L130.1 183.2 L130.5 203.0 L130.8 206.0 L131.2 216.1 L131.5 194.4 L131.9 221.1 L132.2 221.4 L132.6 212.0 L132.9 222.7 L133.2 193.6 L133.6 113.0 L133.9 173.8 L134.3 197.0 L134.6 192.2 L135.0 172.6 L135.3 220.6 L135.7 189.7 L136.0 175.2 L136.4 186.7 L136.7 217.2 L137.1 220.5 L137.4 209.3 L137.8 214.5 L138.1 121.8 L138.5 90.6 L138.8 157.2 L139.2 206.7 L139.5 112.1 L139.9 211.3 L140.2 188.6 L140.6 211.6 L140.9 219.6 L141.3 218.2 L141.6 210.4 L141.9 219.4 L142.3 208.4 L142.6 206.6 L143.0 173.1 L143.3 224.7 L143.7 210.4 L144.0 186.7 L144.4 190.0 L144.7 174.5 L145.1 224.8 L145.4 224.3 L145.8 201.7 L146.1 206.0 L146.5 209.1 L146.8 207.1 L147.2 222.7 L147.5 201.1 L147.9 209.8 L148.2 218.4 L148.6 218.2 L148.9 222.2 L149.3 224.6 L149.6 216.3 L150.0 222.7 L150.3 193.5 L150.6 186.7 L151.0 205.3 L151.3 222.8 L151.7 214.5 L152.0 219.1 L152.4 199.3 L152.7 217.8 L153.1 223.2 L153.4 195.9 L153.8 226.2 L154.1 208.8 L154.5 223.7 L154.8 217.8 L155.2 219.0 L155.5 206.2 L155.9 156.8 L156.2 228.0 L156.6 220.4 L156.9 223.9 L157.3 225.2 L157.6 227.9 L158.0 208.2 L158.3 202.8 L158.7 190.0 L159.0 143.9 L159.3 183.7 L159.7 215.0 L160.0 199.0 L160.4 203.9 L160.7 224.5 L161.1 230.1 L161.4 231.1 L161.8 218.4 L162.1 217.1 L162.5 214.8 L162.8 221.8 L163.2 207.5 L163.5 215.8 L163.9 220.7 L164.2 231.0 L164.6 227.0 L164.9 219.0 L165.3 221.9 L165.6 215.4 L166.0 229.2 L166.3 212.3 L166.7 214.4 L167.0 205.5 L167.4 210.0 L167.7 194.8 L168.0 210.8 L168.4 226.9 L168.7 210.1 L169.1 196.8 L169.4 225.1 L169.8 216.9 L170.1 211.4 L170.5 179.8 L170.8 188.4 L171.2 222.7 L171.5 223.3 L171.9 125.9 L172.2 223.5 L172.6 217.0 L172.9 231.4 L173.3 209.5 L173.6 202.5 L174.0 230.8 L174.3 229.2 L174.7 167.9 L175.0 109.2 L175.4 173.2 L175.7 210.8 L176.1 202.6 L176.4 194.3 L176.7 186.6 L177.1 233.5 L177.4 229.8 L177.8 211.1 L178.1 225.1 L178.5 212.3 L178.8 227.9 L179.2 213.9 L179.5 234.2 L179.9 223.3 L180.2 232.0 L180.6 195.8 L180.9 210.5 L181.3 160.0 L181.6 194.1 L182.0 231.7 L182.3 220.2 L182.7 201.3 L183.0 232.6 L183.4 229.1 L183.7 176.0 L184.1 150.0 L184.4 232.5 L184.8 232.3 L185.1 229.4 L185.4 233.8 L185.8 232.4 L186.1 222.7 L186.5 228.7 L186.8 232.0 L187.2 216.8 L187.5 223.9 L187.9 224.8 L188.2 220.3 L188.6 212.8 L188.9 225.6 L189.3 234.7 L189.6 231.2 L190.0 195.5 L190.3 226.1 L190.7 231.3 L191.0 228.5 L191.4 188.6 L191.7 220.6 L192.1 220.6 L192.4 232.7 L192.8 231.6 L193.1 231.1 L193.5 172.9 L193.8 220.2 L194.1 229.5 L194.5 222.4 L194.8 210.4 L195.2 232.3 L195.5 215.1 L195.9 207.9 L196.2 225.4 L196.6 217.6 L196.9 235.3 L197.3 205.2 L197.6 202.6 L198.0 127.1 L198.3 212.0 L198.7 202.2 L199.0 224.7 L199.4 232.3 L199.7 221.5 L200.1 232.3 L200.4 218.3 L200.8 227.7 L201.1 231.6 L201.5 235.9 L201.8 222.8 L202.2 229.2 L202.5 223.2 L202.8 226.3 L203.2 230.2 L203.5 228.4 L203.9 222.6 L204.2 225.6 L204.6 234.0 L204.9 224.9 L205.3 236.3 L205.6 236.0 L206.0 237.7 L206.3 231.2 L206.7 226.8 L207.0 180.2 L207.4 97.5 L207.7 176.9 L208.1 207.0 L208.4 211.0 L208.8 219.2 L209.1 212.8 L209.5 177.4 L209.8 219.0 L210.2 229.3 L210.5 198.5 L210.9 175.5 L211.2 178.7 L211.5 215.4 L211.9 234.6 L212.2 232.7 L212.6 230.8 L212.9 236.0 L213.3 228.5 L213.6 231.7 L214.0 238.9 L214.3 238.3 L214.7 240.0 L215.0 240.0 L215.4 235.6 L215.7 241.3 L216.1 241.0 L216.4 238.2 L216.8 191.7 L217.1 222.2 L217.5 228.0 L217.8 241.0 L218.2 226.5 L218.5 183.2 L218.9 231.3 L219.2 235.2 L219.6 190.6 L219.9 226.6 L220.2 227.8 L220.6 237.3 L220.9 219.3 L221.3 195.6 L221.6 213.8 L222.0 234.0 L222.3 231.0 L222.7 216.3 L223.0 208.8 L223.4 220.1 L223.7 189.4 L224.1 220.7 L224.4 228.4 L224.8 236.4 L225.1 236.9 L225.5 233.8 L225.8 240.3 L226.2 208.1 L226.5 219.1 L226.9 178.2 L227.2 231.2 L227.6 211.5 L227.9 165.9 L228.3 242.0 L228.6 241.0 L228.9 240.5 L229.3 233.6 L229.6 232.3 L230.0 214.5 L230.3 229.4 L230.7 152.2 L231.0 202.6 L231.4 174.6 L231.7 73.4 L232.1 229.6 L232.4 241.7 L232.8 233.7 L233.1 210.0 L233.5 173.1 L233.8 237.8 L234.2 238.8 L234.5 215.9 L234.9 228.4 L235.2 242.3 L235.6 232.9 L235.9 193.2 L236.3 205.5 L236.6 213.5 L237.0 240.1 L237.3 241.4 L237.6 234.7 L238.0 199.3 L238.3 226.0 L238.7 202.3 L239.0 243.1 L239.4 238.6 L239.7 238.8 L240.1 219.3 L240.4 213.0 L240.8 179.3 L241.1 223.6 L241.5 235.2 L241.8 239.8 L242.2 240.3 L242.5 173.7 L242.9 241.5 L243.2 239.3 L243.6 230.7 L243.9 233.5 L244.3 243.3 L244.6 235.6 L245.0 192.6 L245.3 172.1 L245.7 169.9 L246.0 237.2 L246.3 183.7 L246.7 231.7 L247.0 237.1 L247.4 233.8 L247.7 238.8 L248.1 202.5 L248.4 236.3 L248.8 223.5 L249.1 233.6 L249.5 246.1 L249.8 230.7 L250.2 225.7 L250.5 202.8 L250.9 190.9 L251.2 219.1 L251.6 239.8 L251.9 245.3 L252.3 237.8 L252.6 199.8 L253.0 240.8 L253.3 232.1 L253.7 226.9 L254.0 222.1 L254.4 236.7 L254.7 243.8 L255.0 189.3 L255.4 213.4 L255.7 80.6 L256.1 221.8 L256.4 234.0 L256.8 223.2 L257.1 206.4 L257.5 229.1 L257.8 243.0 L258.2 225.9 L258.5 99.7 L258.9 224.0 L259.2 201.7 L259.6 237.3 L259.9 209.6 L260.3 240.0 L260.6 222.9 L261.0 211.1 L261.3 218.0 L261.7 245.9 L262.0 237.4 L262.4 244.1 L262.7 245.2 L263.1 248.0 L263.4 239.8 L263.7 229.5 L264.1 201.9 L264.4 239.5 L264.8 240.9 L265.1 237.2 L265.5 244.1 L265.8 224.3 L266.2 243.5 L266.5 236.2 L266.9 229.0 L267.2 225.4 L267.6 247.9 L267.9 230.9 L268.3 200.6 L268.6 244.3 L269.0 248.2 L269.3 248.4 L269.7 249.1 L270.0 234.5 L270.4 181.5 L270.7 237.0 L271.1 246.3 L271.4 235.0 L271.8 248.7 L272.1 206.2 L272.4 239.6 L272.8 234.8 L273.1 250.1 L273.5 236.3 L273.8 235.7 L274.2 240.5 L274.5 247.3 L274.9 246.6 L275.2 218.8 L275.6 236.2 L275.9 246.2 L276.3 236.6 L276.6 219.6 L277.0 239.7 L277.3 226.1 L277.7 241.5 L278.0 248.7 L278.4 248.4 L278.7 243.1 L279.1 248.6 L279.4 250.7 L279.8 238.4 L280.1 187.7 L280.5 239.4 L280.8 242.6 L281.1 249.1 L281.5 250.0 L281.8 240.4 L282.2 247.5 L282.5 228.4 L282.9 229.4 L283.2 214.2 L283.6 229.8 L283.9 248.1 L284.3 248.4 L284.6 212.4 L285.0 212.5 L285.3 212.0 L285.7 244.6 L286.0 240.0 L286.4 238.3 L286.7 230.7 L287.1 245.1 L287.4 241.6 L287.8 240.9 L288.1 227.4 L288.5 247.4 L288.8 221.0 L289.2 234.2 L289.5 242.9 L289.8 248.6 L290.2 247.2 L290.5 249.1 L290.9 249.0 L291.2 251.2 L291.6 235.3 L291.9 162.4 L292.3 246.9 L292.6 243.5 L293.0 219.4 L293.3 251.8 L293.7 228.5 L294.0 239.6 L294.4 246.7 L294.7 245.8 L295.1 250.7 L295.4 227.4 L295.8 245.3 L296.1 161.2 L296.5 239.5 L296.8 240.9 L297.2 246.6 L297.5 238.3 L297.9 243.5 L298.2 246.3 L298.5 248.9 L298.9 239.7 L299.2 246.4 L299.6 244.1 L299.9 198.9 L300.3 240.2 L300.6 236.2 L301.0 234.2 L301.3 244.1 L301.7 244.8 L302.0 246.5 L302.4 243.1 L302.7 228.8 L303.1 226.0 L303.4 246.7 L303.8 248.6 L304.1 247.8 L304.5 242.2 L304.8 243.0 L305.2 252.8 L305.5 246.8 L305.9 240.2 L306.2 240.0 L306.6 143.2 L306.9 214.0 L307.2 236.6 L307.6 244.2 L307.9 246.5 L308.3 254.5 L308.6 255.3 L309.0 240.1 L309.3 249.1 L309.7 250.3 L310.0 233.4 L310.4 248.7 L310.7 232.2 L311.1 215.9 L311.4 253.1 L311.8 254.6 L312.1 237.7 L312.5 234.7 L312.8 242.0 L313.2 237.4 L313.5 202.9 L313.9 227.9 L314.2 181.2 L314.6 228.2 L314.9 244.3 L315.3 249.3 L315.6 246.8 L315.9 245.9 L316.3 247.2 L316.6 208.2 L317.0 231.7 L317.3 223.2 L317.7 196.9 L318.0 246.5 L318.4 249.6 L318.7 253.2 L319.1 244.3 L319.4 234.0 L319.8 244.4 L320.1 241.1 L320.5 255.2 L320.8 235.6 L321.2 233.6 L321.5 248.5 L321.9 217.9 L322.2 256.2 L322.6 252.7 L322.9 244.3 L323.3 249.1 L323.6 222.7 L324.0 249.6 L324.3 243.2 L324.6 246.9 L325.0 230.9 L325.3 226.3 L325.7 206.9 L326.0 225.9 L326.4 256.0 L326.7 256.3 L327.1 232.6 L327.4 238.9 L327.8 243.9 L328.1 250.3 L328.5 229.9 L328.8 251.9 L329.2 250.7 L329.5 226.7 L329.9 246.2 L330.2 221.9 L330.6 243.2 L330.9 148.0 L331.3 167.8 L331.6 227.2 L332.0 234.1 L332.3 236.9 L332.7 230.2 L333.0 207.8 L333.3 228.2 L333.7 252.5 L334.0 242.1 L334.4 253.1 L334.7 249.3 L335.1 244.8 L335.4 142.9 L335.8 244.4 L336.1 227.0 L336.5 250.8 L336.8 149.3 L337.2 233.3 L337.5 229.0 L337.9 243.1 L338.2 224.9 L338.6 251.1 L338.9 253.3 L339.3 253.9 L339.6 226.0 L340.0 254.1 L340.3 251.1 L340.7 246.0 L341.0 258.6 L341.4 242.5 L341.7 255.8 L342.0 258.1 L342.4 259.5 L342.7 243.2 L343.1 197.2 L343.4 248.2 L343.8 236.4 L344.1 244.0 L344.5 243.5 L344.8 238.4 L345.2 202.0 L345.5 232.6 L345.9 212.6 L346.2 251.4 L346.6 247.5 L346.9 231.2 L347.3 231.1 L347.6 250.8 L348.0 243.8 L348.3 250.6 L348.7 249.8 L349.0 244.0 L349.4 252.6 L349.7 246.5 L350.1 250.0 L350.4 256.1 L350.7 246.5 L351.1 258.7 L351.4 253.9 L351.8 244.7 L352.1 254.7 L352.5 258.2 L352.8 258.0 L353.2 258.6 L353.5 247.7 L353.9 250.8 L354.2 231.0 L354.6 212.5 L354.9 238.6 L355.3 248.6 L355.6 254.1 L356.0 257.3 L356.3 228.0 L356.7 253.9 L357.0 252.6 L357.4 258.8 L357.7 251.2 L358.1 183.2 L358.4 214.4 L358.8 220.8 L359.1 241.5 L359.4 259.8 L359.8 260.7 L360.1 259.7 L360.5 258.2 L360.8 227.8 L361.2 247.1 L361.5 256.4 L361.9 248.6 L362.2 255.4 L362.6 255.5 L362.9 258.2 L363.3 240.1 L363.6 245.4 L364.0 229.0 L364.3 248.3 L364.7 241.6 L365.0 242.1 L365.4 254.5 L365.7 251.8 L366.1 236.8 L366.4 257.7 L366.8 238.4 L367.1 225.4 L367.5 233.2 L367.8 236.2 L368.1 258.8 L368.5 261.8 L368.8 253.3 L369.2 248.5 L369.5 251.7 L369.9 252.3 L370.2 258.4 L370.6 256.2 L370.9 254.5 L371.3 228.1 L371.6 252.9 L372.0 236.0 L372.3 262.0 L372.7 259.2 L373.0 258.3 L373.4 242.1 L373.7 251.2 L374.1 262.4 L374.4 257.0 L374.8 245.4 L375.1 248.9 L375.5 258.2 L375.8 258.8 L376.2 251.8 L376.5 262.6 L376.8 216.7 L377.2 229.5 L377.5 262.4 L377.9 262.3 L378.2 263.8 L378.6 262.5 L378.9 252.1 L379.3 260.6 L379.6 256.8 L380.0 258.7 L380.3 250.3 L380.7 169.6 L381.0 237.6 L381.4 254.5 L381.7 262.5 L382.1 242.3 L382.4 260.6 L382.8 264.4 L383.1 213.0 L383.5 258.7 L383.8 262.0 L384.2 258.9 L384.5 243.9 L384.9 247.4 L385.2 218.7 L385.5 258.9 L385.9 259.4 L386.2 247.2 L386.6 263.6 L386.9 229.1 L387.3 226.2 L387.6 252.2 L388.0 246.9 L388.3 249.4 L388.7 252.5 L389.0 249.2 L389.4 242.5 L389.7 240.6 L390.1 240.4 L390.4 264.0 L390.8 263.8 L391.1 246.6 L391.5 249.2 L391.8 264.5 L392.2 258.0 L392.5 264.1 L392.9 258.3 L393.2 260.3 L393.6 264.1 L393.9 256.0 L394.2 262.5 L394.6 262.5 L394.9 265.2 L395.3 266.0 L395.6 250.4 L396.0 253.6 L396.3 257.6 L396.7 264.2 L397.0 263.5 L397.4 260.7 L397.7 261.9 L398.1 260.9 L398.4 253.2 L398.8 261.2 L399.1 263.1 L399.5 240.8 L399.8 260.9 L400.2 260.4 L400.5 263.9 L400.9 261.5 L401.2 267.3 L401.6 253.5 L401.9 251.9 L402.3 230.4 L402.6 241.6 L402.9 254.5 L403.3 262.4 L403.6 256.1 L404.0 253.6 L404.3 251.1 L404.7 251.7 L405.0 258.1 L405.4 260.4 L405.7 267.4 L406.1 254.7 L406.4 260.1 L406.8 256.0 L407.1 262.6 L407.5 254.0 L407.8 263.9 L408.2 267.2 L408.5 265.5 L408.9 216.1 L409.2 226.5 L409.6 234.4 L409.9 253.7 L410.3 238.6 L410.6 256.1 L411.0 265.9 L411.3 260.1 L411.6 267.6 L412.0 268.8 L412.3 254.5 L412.7 264.2 L413.0 262.5 L413.4 268.0 L413.7 259.6 L414.1 262.2 L414.4 214.9 L414.8 260.6 L415.1 267.9 L415.5 254.2 L415.8 247.9 L416.2 187.9 L416.5 260.1 L416.9 257.4 L417.2 260.9 L417.6 261.8 L417.9 268.4 L418.3 256.3 L418.6 262.2 L419.0 257.3 L419.3 268.4 L419.7 266.7 L420.0 266.1 L420.3 258.3 L420.7 260.3 L421.0 232.3 L421.4 231.7 L421.7 236.4 L422.1 251.4 L422.4 262.6 L422.8 257.3 L423.1 267.6 L423.5 260.4 L423.8 227.9 L424.2 255.5 L424.5 245.5 L424.9 256.0 L425.2 261.4 L425.6 261.2 L425.9 264.2 L426.3 266.3 L426.6 261.0 L427.0 270.7 L427.3 254.1 L427.7 240.2 L428.0 265.8 L428.4 270.3 L428.7 268.1 L429.0 266.5 L429.4 245.9 L429.7 142.1 L430.1 264.4 L430.4 250.4 L430.8 269.4 L431.1 253.2 L431.5 151.4 L431.8 253.7 L432.2 270.6 L432.5 253.5 L432.9 270.2 L433.2 241.8 L433.6 258.0 L433.9 270.4 L434.3 248.3 L434.6 182.9 L435.0 186.4 L435.3 250.4 L435.7 255.8 L436.0 256.8 L436.4 270.6 L436.7 269.7 L437.1 270.3 L437.4 271.6 L437.7 269.4 L438.1 265.5 L438.4 271.5 L438.8 265.5 L439.1 270.4 L439.5 242.9 L439.8 262.4 L440.2 212.2 L440.5 255.9 L440.9 267.0 L441.2 269.8 L441.6 263.5 L441.9 271.2 L442.3 254.6 L442.6 265.3 L443.0 258.5 L443.3 253.8 L443.7 272.0 L444.0 269.6 L444.4 262.6 L444.7 271.6 L445.1 272.5 L445.4 270.1 L445.8 247.8 L446.1 258.3 L446.4 265.2 L446.8 232.1 L447.1 243.3 L447.5 270.0 L447.8 272.5 L448.2 272.4 L448.5 258.3 L448.9 262.2 L449.2 273.7 L449.6 253.3 L449.9 263.5 L450.3 262.3 L450.6 249.0 L451.0 253.6 L451.3 272.2 L451.7 235.7 L452.0 257.3 L452.4 255.1 L452.7 224.9 L453.1 265.0 L453.4 269.1 L453.8 274.4 L454.1 274.8 L454.5 255.2 L454.8 261.8 L455.1 268.1 L455.5 269.1 L455.8 243.3 L456.2 259.8 L456.5 243.6 L456.9 271.4 L457.2 266.2 L457.6 272.0 L457.9 273.8 L458.3 273.7 L458.6 267.7 L459.0 229.8 L459.3 270.2 L459.7 264.2 L460.0 272.7 L460.4 267.2 L460.7 261.9 L461.1 182.1 L461.4 274.1 L461.8 275.9 L462.1 269.7 L462.5 270.6 L462.8 272.3 L463.2 274.5 L463.5 272.5 L463.8 275.9 L464.2 269.0 L464.5 275.9 L464.9 271.9 L465.2 273.8 L465.6 274.4 L465.9 271.4 L466.3 243.5 L466.6 228.8 L467.0 251.0 L467.3 273.6 L467.7 268.6 L468.0 249.4 L468.4 261.4 L468.7 268.7 L469.1 269.5 L469.4 256.2 L469.8 261.2 L470.1 261.6 L470.5 270.6 L470.8 262.4 L471.2 248.9 L471.5 262.8 L471.9 273.4 L472.2 273.6 L472.5 274.1 L472.9 266.4 L473.2 277.2 L473.6 257.3 L473.9 269.2 L474.3 276.0 L474.6 278.0 L475.0 269.1 L475.3 276.7 L475.7 276.0 L476.0 270.5 L476.4 267.4 L476.7 271.6 L477.1 276.4 L477.4 276.7 L477.8 276.4 L478.1 278.9 L478.5 272.1 L478.8 277.3 L479.2 270.9 L479.5 279.0 L479.9 278.8 L480.2 277.0 L480.6 278.2 L480.9 275.5 L481.2 263.3 L481.6 279.6 L481.9 278.8 L482.3 276.9 L482.6 274.0 L483.0 279.8 L483.3 274.7 L483.7 243.7 L484.0 271.0 L484.4 166.1 L484.7 244.3 L485.1 265.4 L485.4 261.0 L485.8 277.3 L486.1 270.3 L486.5 256.6 L486.8 265.6 L487.2 274.0 L487.5 275.5 L487.9 277.3 L488.2 271.7 L488.6 269.9 L488.9 276.8 L489.3 277.0 L489.6 279.8 L489.9 280.1 L490.3 276.1 L490.6 272.2 L491.0 269.0 L491.3 271.5 L491.7 280.1 L492.0 280.9 L492.4 279.8 L492.7 267.7 L493.1 277.4 L493.4 280.1 L493.8 281.0 L494.1 279.2 L494.5 280.5 L494.8 279.5 L495.2 277.6 L495.5 272.9 L495.9 275.7 L496.2 276.1 L496.6 278.0 L496.9 279.5 L497.3 281.5 L497.6 281.7 L498.0 281.5 L498.3 274.8 L498.6 278.4 L499.0 277.5 L499.3 280.0 L499.7 279.3 L500.0 276.0 L500.4 279.5 L500.7 281.1 L501.1 273.2 L501.4 274.7 L501.8 280.3 L502.1 278.9 L502.5 282.4 L502.8 280.9 L503.2 281.7 L503.5 281.0 L503.9 282.7 L504.2 279.1 L504.6 281.6 L504.9 282.5 L505.3 274.3 L505.6 228.3 L506.0 238.7 L506.3 277.0 L506.7 277.1 L507.0 274.2 L507.3 272.4 L507.7 279.0 L508.0 282.9 L508.4 282.6 L508.7 281.4 L509.1 282.7 L509.4 273.6 L509.8 271.8 L510.1 280.9 L510.5 281.6 L510.8 276.2 L511.2 281.4 L511.5 282.2 L511.9 281.5 L512.2 283.4 L512.6 283.2 L512.9 283.6 L513.3 282.2 L513.6 282.7 L514.0 284.1 L514.3 283.1 L514.7 280.4 L515.0 281.9 L515.4 283.2 L515.7 276.5 L516.0 262.4 L516.4 249.2 L516.7 281.2 L517.1 282.2 L517.4 281.3 L517.8 280.5 L518.1 282.9 L518.5 283.7 L518.8 283.8 L519.2 281.1 L519.5 284.0 L519.9 281.5 L520.2 280.4 L520.6 282.9 L520.9 281.9 L521.3 280.4 L521.6 284.5 L522.0 283.2 L522.3 282.8 L522.7 276.5 L523.0 284.4 L523.4 275.4 L523.7 279.8 L524.1 282.4 L524.4 282.8 L524.7 284.5 L525.1 279.4 L525.4 284.6 L525.8 282.6 L526.1 281.1 L526.5 278.9 L526.8 284.3 L527.2 283.6 L527.5 284.0 L527.9 284.5 L528.2 277.7 L528.6 285.1 L528.9 284.7 L529.3 283.5 L529.6 281.4 L530.0 281.7 L530.3 284.6 L530.7 284.7 L531.0 284.8 L531.4 285.6 L531.7 285.5 L532.1 283.8 L532.4 285.9 L532.8 285.5 L533.1 285.5 L533.4 284.2 L533.8 282.9 L534.1 284.5 L534.5 285.7 L534.8 286.2 L535.2 281.3 L535.5 281.7 L535.9 283.9 L536.2 259.8 L536.6 248.4 L536.9 240.0 L537.3 181.7 L537.6 281.8 L538.0 284.7 L538.3 285.5 L538.7 285.3 L539.0 285.9 L539.4 284.7 L539.7 286.5 L540.1 286.6 L540.4 286.6 L540.8 284.7 L541.1 286.5 L541.5 286.5 L541.8 285.9 L542.1 286.9 L542.5 285.5 L542.8 286.9 L543.2 286.9 L543.5 286.3 L543.9 285.9 L544.2 284.9 L544.6 285.6 L544.9 282.8 L545.3 285.1 L545.6 287.1 L546.0 287.1 L546.3 286.4 L546.7 286.5 L547.0 284.9 L547.4 286.6 L547.7 277.6 L548.1 281.4 L548.4 279.8 L548.8 287.4 L549.1 287.8 L549.5 286.9 L549.8 285.9 L550.2 286.6 L550.5 284.6 L550.8 282.8 L551.2 287.1 L551.5 288.0 L551.9 287.4 L552.2 283.2 L552.6 287.3 L552.9 288.0 L553.3 288.2 L553.6 283.6 L554.0 286.6 L554.3 287.2 L554.7 285.4 L555.0 288.3 L555.4 286.7 L555.7 283.9 L556.1 286.4 L556.4 288.2 L556.8 288.2 L557.1 286.6 L557.5 288.4 L557.8 287.4 L558.2 286.3 L558.5 286.4 L558.9 288.2 L559.2 287.4 L559.5 288.1 L559.9 288.2 L560.2 286.7 L560.6 288.5 L560.9 288.6 L561.3 288.2 L561.6 288.8 L562.0 288.6 L562.3 288.4 L562.7 288.2 L563.0 288.3 L563.4 288.7 L563.7 287.9 L564.1 288.5 L564.4 282.3 L564.8 288.7 L565.1 287.6 L565.5 286.4 L565.8 287.8 L566.2 289.0 L566.5 289.1 L566.9 288.8 L567.2 288.4 L567.6 285.3 L567.9 287.3 L568.2 287.9 L568.6 285.8 L568.9 280.3 L569.3 278.8 L569.6 281.1 L570.0 281.1 L570.3 285.3 L570.7 286.6 L571.0 285.3 L571.4 289.0 L571.7 288.3 L572.1 288.9 L572.4 289.2 L572.8 288.8 L573.1 288.6 L573.5 289.2 L573.8 289.4 L574.2 288.7 L574.5 288.6 L574.9 288.4 L575.2 287.6 L575.6 288.9 L575.9 288.8 L576.3 287.7 L576.6 288.6 L576.9 289.3 L577.3 288.0 L577.6 288.2 L578.0 289.3 L578.3 289.4 L578.7 288.9 L579.0 288.3 L579.4 289.1 L579.7 289.8 L580.1 289.5 L580.4 290.0 L580.8 289.6 L581.1 290.2 L581.5 289.6 L581.8 290.0 L582.2 289.5 L582.5 290.2 L582.9 289.4 L583.2 290.3 L583.6 290.0 L583.9 290.3 L584.3 290.1 L584.6 288.6 L585.0 288.6 L585.3 289.7 L585.6 290.0 L586.0 290.4 L586.3 289.9 L586.7 289.9 L587.0 290.5 L587.4 289.7 L587.7 290.0 L588.1 290.6 L588.4 290.4 L588.8 290.6 L589.1 290.1 L589.5 290.6 L589.8 289.6 L590.2 290.6 L590.5 290.7 L590.9 290.7 L591.2 290.4 L591.6 290.7 L591.9 290.3 L592.3 290.6 L592.6 290.8 L593.0 288.6 L593.3 290.0 L593.7 289.9 L594.0 289.1 L594.3 289.8 L594.7 290.4 L595.0 289.8 L595.4 290.3 L595.7 291.0 L596.1 290.8 L596.4 289.0 L596.8 290.4 L597.1 290.0 L597.5 290.9 L597.8 290.9 L598.2 291.0 L598.5 291.0 L598.9 289.7 L599.2 289.0 L599.6 291.1 L599.9 290.1 L600.3 289.9 L600.6 291.1 L601.0 291.3 L601.3 290.7 L601.7 291.3 L602.0 291.3 L602.4 290.9 L602.7 291.3 L603.0 291.3 L603.4 291.1 L603.7 291.5 L604.1 291.2 L604.4 290.6 L604.8 291.5 L605.1 291.2 L605.5 291.2 L605.8 291.0 L606.2 291.0 L606.5 290.9 L606.9 291.1 L607.2 290.9 L607.6 290.8 L607.9 291.7 L608.3 289.5 L608.6 288.9 L609.0 290.1 L609.3 291.3 L609.7 291.6 L610.0 290.3 L610.4 289.6 L610.7 291.5 L611.1 291.7 L611.4 291.9 L611.7 291.5 L612.1 291.5 L612.4 290.8 L612.8 291.5 L613.1 291.9 L613.5 291.7 L613.8 291.0 L614.2 289.5 L614.5 290.8 L614.9 291.6 L615.2 291.9 L615.6 291.8 L615.9 291.9 L616.3 291.8 L616.6 291.1 L617.0 290.7 L617.3 291.6 L617.7 291.7 L618.0 292.0 L618.4 292.0 L618.7 291.8 L619.1 292.1 L619.4 292.0 L619.8 292.1 L620.1 292.0 L620.4 292.2 L620.8 291.9 L621.1 292.1 L621.5 292.1 L621.8 291.7 L622.2 291.3 L622.5 292.1 L622.9 292.1 L623.2 292.4 L623.6 291.7 L623.9 292.2 L624.3 292.4 L624.6 291.5 L625.0 292.3 L625.3 292.5 L625.7 291.5 L626.0 292.1 L626.4 292.4 L626.7 292.4 L627.1 292.3 L627.4 292.3 L627.8 291.1 L628.1 292.0 L628.5 292.2 L628.8 292.4 L629.1 291.8 L629.5 292.2 L629.8 292.2 L630.2 291.5 L630.5 290.5 L630.9 291.8 L631.2 291.7 L631.6 291.3 L631.9 292.5 L632.3 292.1 L632.6 292.3 L633.0 291.8 L633.3 291.6 L633.7 292.5 L634.0 292.4 L634.4 292.2 L634.7 292.1 L635.1 291.9 L635.4 292.8 L635.8 292.8 L636.1 292.7 L636.5 292.9 L636.8 293.0 L637.2 292.9 L637.5 291.9 L637.8 292.9 L638.2 292.7 L638.5 292.7 L638.9 292.9 L639.2 291.4 L639.6 292.5 L639.9 292.8 L640.3 292.4 L640.6 292.8 L641.0 292.9 L641.3 292.8 L641.7 293.1 L642.0 292.9 L642.4 293.0 L642.7 292.8 L643.1 291.7 L643.4 293.2 L643.8 292.7 L644.1 292.7 L644.5 293.2 L644.8 293.2 L645.2 293.0 L645.5 293.0 L645.9 292.8 L646.2 293.1 L646.5 292.7 L646.9 293.1 L647.2 293.2 L647.6 293.3 L647.9 293.3 L648.3 293.0 L648.6 293.3 L649.0 293.3 L649.3 293.2 L649.7 292.2 L650.0 292.4 L650.4 293.3 L650.7 293.4 L651.1 293.4 L651.4 293.3 L651.8 293.4 L652.1 293.1 L652.5 292.2 L652.8 292.6 L653.2 293.3 L653.5 292.5 L653.9 292.8 L654.2 293.5 L654.6 292.2 L654.9 292.9 L655.2 293.0 L655.6 292.6 L655.9 293.3 L656.3 293.4 L656.6 293.3 L657.0 293.5 L657.3 292.9 L657.7 293.4 L658.0 293.7 L658.4 293.6 L658.7 293.4 L659.1 293.7 L659.4 293.2 L659.8 293.4 L660.1 293.8 L660.5 293.7 L660.8 293.8 L661.2 293.7 L661.5 293.7 L661.9 293.8 L662.2 293.7 L662.6 293.7 L662.9 293.7 L663.3 293.7 L663.6 293.2 L663.9 292.2 L664.3 293.2 L664.6 293.8 L665.0 293.9 L665.3 293.7 L665.7 293.8 L666.0 293.8 L666.4 293.6 L666.7 293.3 L667.1 293.9 L667.4 293.8 L667.8 293.5 L668.1 293.9 L668.5 293.4 L668.8 293.8 L669.2 294.0 L669.5 293.9 L669.9 294.0 L670.2 293.9 L670.6 293.1 L670.9 293.7 L671.3 293.9 L671.6 293.9 L672.0 294.1 L672.3 293.8 L672.6 294.1 L673.0 294.1 L673.3 293.6 L673.7 294.1 L674.0 294.1 L674.4 294.2 L674.7 294.2 L675.1 294.2 L675.4 294.2 L675.8 293.9 L676.1 294.2 L676.5 293.9 L676.8 294.2 L677.2 294.1 L677.5 294.0 L677.9 294.2 L678.2 294.0 L678.6 293.7 L678.9 293.9 L679.3 294.3 L679.6 294.3 L680.0 294.3 L680.3 293.9 L680.7 294.3 L681.0 294.0 L681.3 294.3 L681.7 294.3 L682.0 294.4 L682.4 294.0 L682.7 294.2 L683.1 294.1 L683.4 294.4 L683.8 294.4 L684.1 293.7 L684.5 294.3 L684.8 294.4 L685.2 294.3 L685.5 294.5 L685.9 294.4 L686.2 294.4 L686.6 294.4 L686.9 294.5 L687.3 294.3 L687.6 294.2 L688.0 294.4 L688.3 293.1 L688.7 294.1 L689.0 294.5 L689.4 294.4 L689.7 294.4 L690.0 294.2 L690.4 294.6 L690.7 294.5 L691.1 294.4 L691.4 294.3 L691.8 294.6 L692.1 294.5 L692.5 294.4 L692.8 294.4 L693.2 294.6 L693.5 294.7 L693.9 294.7 L694.2 294.3 L694.6 294.2 L694.9 294.7 L695.3 294.6 L695.6 294.7 L696.0 294.7 L696.3 294.7 L696.7 294.6 L697.0 294.5 L697.4 294.6 L697.7 294.2 L698.1 294.8 L698.4 294.6 L698.7 294.6 L699.1 294.8 L699.4 294.8 L699.8 294.1 L700.1 294.0 L700.5 294.7 L700.8 294.7 L701.2 294.8 L701.5 294.8 L701.9 294.9 L702.2 294.8 L702.6 294.8 L702.9 294.6 L703.3 294.8 L703.6 294.9 L704.0 294.8 L704.3 294.9 L704.7 294.9 L705.0 294.9 L705.4 294.8 L705.7 294.8 L706.1 294.6 L706.4 294.8 L706.8 295.0 L707.1 294.9 L707.4 294.7 L707.8 294.7 L708.1 295.0 L708.5 295.1 L708.8 295.0 L709.2 294.9 L709.5 294.9 L709.9 295.0 L710.2 295.1 L710.6 294.9 L710.9 294.2 L711.3 295.0 L711.6 295.1 L712.0 295.1 L712.3 295.1 L712.7 295.1 L713.0 295.1 L713.4 295.1 L713.7 294.8 L714.1 294.8 L714.4 295.1 L714.8 295.2 L715.1 295.2 L715.5 295.1 L715.8 294.8 L716.1 295.2 L716.5 295.2 L716.8 295.1 L717.2 295.2 L717.5 295.2 L717.9 295.2 L718.2 295.2 L718.6 295.3 L718.9 295.1 L719.3 294.9 L719.6 295.0 L720.0 295.2 L720.3 295.2 L720.7 295.1 L721.0 295.3 L721.4 295.2 L721.7 295.3 L722.1 295.3 L722.4 295.2 L722.8 295.3 L723.1 295.3 L723.5 295.2 L723.8 295.3 L724.2 295.2 L724.5 294.7 L724.8 295.4 L725.2 295.3 L725.5 295.3 L725.9 295.3 L726.2 295.4 L726.6 295.2 L726.9 295.4 L727.3 295.4 L727.6 295.4 L728.0 295.1 L728.3 295.4 L728.7 295.4 L729.0 295.5 L729.4 295.5 L729.7 295.1 L730.1 295.1 L730.4 295.4 L730.8 295.2 L731.1 295.5 L731.5 295.5 L731.8 295.5 L732.2 295.5 L732.5 295.4 L732.9 295.5 L733.2 295.3 L733.5 295.2 L733.9 295.3 L734.2 295.6 L734.6 295.4 L734.9 295.5 L735.3 295.5 L735.6 295.5 L736.0 295.6 L736.3 295.3 L736.7 295.5 L737.0 295.6 L737.4 295.6 L737.7 295.5 L738.1 295.6 L738.4 295.5 L738.8 295.7 L739.1 295.6 L739.5 295.6 L739.8 295.6 L740.2 295.7 L740.5 295.7 L740.9 295.7 L741.2 295.7 L741.6 295.7 L741.9 295.5 L742.2 295.6 L742.6 295.5 L742.9 295.6 L743.3 295.5 L743.6 295.5 L744.0 295.6 L744.3 295.5 L744.7 295.6 L745.0 295.7 L745.4 295.8 L745.7 295.4 L746.1 295.8 L746.4 295.8 L746.8 295.8 L747.1 295.8 L747.5 295.7 L747.8 295.8 L748.2 295.8 L748.5 295.8 L748.9 295.8 L749.2 295.3 L749.6 295.7 L749.9 295.8 L750.3 295.8 L750.6 295.9 L750.9 295.9 L751.3 295.9 L751.6 295.9 L752.0 295.9 L752.3 295.8 L752.7 295.8 L753.0 295.7 L753.4 295.7 L753.7 295.8 L754.1 295.8 L754.4 295.9 L754.8 295.9 L755.1 295.9 L755.5 295.9 L755.8 295.9 L756.2 295.9 L756.5 295.9 L756.9 295.9 L757.2 295.9 L757.6 296.0 L757.9 295.8 L758.3 295.9 L758.6 295.8 L759.0 295.9 L759.3 296.0 L759.6 296.0 L760.0 295.9 L760.3 296.0 L760.7 296.0 L761.0 295.9 L761.4 295.8 L761.7 296.0 L762.1 295.9 L762.4 296.0 L762.8 296.0 L763.1 296.0 L763.5 296.0 L763.8 296.0 L764.2 296.0 L764.5 296.0 L764.9 295.9 L765.2 296.1 L765.6 296.1 L765.9 296.1 L766.3 296.1 L766.6 296.1 L767.0 296.1 L767.3 296.1 L767.7 296.1 L768.0 296.1"/>
<path class="dd-edge-acc" style="stroke-width:2.4;fill:none" d="M72.0 133.7 L75.5 134.7 L79.0 142.0 L82.4 147.2 L85.9 156.1 L89.4 162.0 L92.9 169.8 L96.4 172.9 L99.8 175.0 L103.3 178.7 L106.8 183.2 L110.3 186.0 L113.8 189.6 L117.2 192.9 L120.7 195.2 L124.2 196.9 L127.7 195.8 L131.2 190.4 L134.6 194.3 L138.1 194.1 L141.6 197.4 L145.1 202.0 L148.6 209.8 L152.0 208.5 L155.5 211.5 L159.0 211.9 L162.5 211.6 L166.0 211.3 L169.4 209.7 L172.9 208.3 L176.4 206.6 L179.9 210.7 L183.4 212.3 L186.8 218.0 L190.3 216.3 L193.8 219.8 L197.3 220.3 L200.8 216.4 L204.2 214.1 L207.7 220.6 L211.2 218.8 L214.7 217.2 L218.2 221.9 L221.6 223.9 L225.1 214.9 L228.6 216.1 L232.1 217.0 L235.6 217.1 L239.0 216.0 L242.5 222.2 L246.0 223.2 L249.5 219.8 L253.0 216.9 L256.4 221.6 L259.9 223.3 L263.4 224.8 L266.9 230.2 L270.4 236.1 L273.8 236.6 L277.3 235.8 L280.8 235.7 L284.3 236.1 L287.8 234.7 L291.2 235.5 L294.7 236.7 L298.2 236.4 L301.7 237.4 L305.2 238.2 L308.6 235.8 L312.1 236.7 L315.6 237.8 L319.1 236.6 L322.6 235.7 L326.0 235.1 L329.5 231.6 L333.0 232.6 L336.5 233.8 L340.0 234.7 L343.4 239.1 L346.9 244.0 L350.4 242.3 L353.9 244.1 L357.4 246.4 L360.8 245.3 L364.3 246.3 L367.8 249.5 L371.3 249.8 L374.8 248.8 L378.2 250.1 L381.7 248.5 L385.2 249.7 L388.7 251.1 L392.2 254.6 L395.6 254.9 L399.1 257.8 L402.6 255.7 L406.1 255.3 L409.6 254.1 L413.0 254.7 L416.5 253.6 L420.0 256.1 L423.5 251.8 L427.0 249.9 L430.4 252.0 L433.9 253.0 L437.4 253.6 L440.9 258.4 L444.4 260.3 L447.8 259.8 L451.3 261.0 L454.8 260.6 L458.3 261.6 L461.8 263.4 L465.2 264.0 L468.7 265.9 L472.2 268.8 L475.7 270.0 L479.2 268.7 L482.6 270.4 L486.1 271.3 L489.6 271.7 L493.1 272.9 L496.6 278.0 L500.0 276.7 L503.5 277.0 L507.0 277.9 L510.5 277.7 L514.0 277.9 L517.4 280.4 L520.9 281.1 L524.4 281.4 L527.9 282.4 L531.4 279.3 L534.8 280.3 L538.3 280.9 L541.8 280.9 L545.3 281.8 L548.8 285.9 L552.2 286.1 L555.7 286.7 L559.2 287.2 L562.7 286.8 L566.2 287.2 L569.6 287.4 L573.1 287.6 L576.6 288.1 L580.1 289.3 L583.6 289.6 L587.0 289.9 L590.5 290.1 L594.0 290.3 L597.5 290.5 L601.0 290.6 L604.4 290.8 L607.9 291.0 L611.4 291.2 L614.9 291.4 L618.4 291.7 L621.8 291.8 L625.3 291.9 L628.8 292.1 L632.3 292.2 L635.8 292.3 L639.2 292.5 L642.7 292.8 L646.2 292.9 L649.7 293.0 L653.2 293.2 L656.6 293.3 L660.1 293.4 L663.6 293.6 L667.1 293.7 L670.6 293.8 L674.0 294.0 L677.5 294.1 L681.0 294.2 L684.5 294.2 L688.0 294.3 L691.4 294.4 L694.9 294.5 L698.4 294.6 L701.9 294.7 L705.4 294.8 L708.8 294.9 L712.3 295.0 L715.8 295.1 L719.3 295.1 L722.8 295.2 L726.2 295.3 L729.7 295.4 L733.2 295.4 L736.7 295.5 L740.2 295.6 L743.6 295.6 L747.1 295.7 L750.6 295.8 L754.1 295.8 L757.6 295.9 L761.0 295.9 L764.5 296.0 L768.0 296.0"/>
<text class="t-say" x="440.9" y="61.7">细线：每走一步后的真实损失。前 1000 步里</text>
<text class="t-say" x="440.9" y="81.7">有 <tspan class="t-acc">494</tspan> 步比上一步更差。</text>
<text class="t-say t-acc" x="440.9" y="109.7">粗线：前后 50 步的平均，一路往下。</text>
</svg>
</div>
<figcaption>图 3 · 每一步都只看一小部分题，近一半的步子让整体变差，但大方向一直往下</figcaption>
</figure>

细线上下乱跳。前 1000 步里，有 494 步走完以后整体损失比走之前还高；一共 10000 步里，有 3774 步是在“改坏”。原因不难理解：每一步的评审只看了 20 道题，照着这 20 道题改，对这 20 道题有利，对其他 180 道题可能不利。可是把前后几十步平均起来看，粗线一路往下，最后整体损失从 0.70 降到 0.0012，200 个点全部判断正确。

这和人改方案其实是一回事：**单独一轮的改动可能是错的，但只要评审不断，下一轮会把它纠回来。** 训练的不同之处只是一轮极便宜，改坏了也就是几毫秒的事，所以它敢每轮只改一点点、改上几万轮，而不必像人那样每轮都想清楚再动手。

可是，每一步只看脚下的坡度，就像在雾里下山。**会不会走进一个小坑，四面都是上坡，再也出不来？**

## 二、会不会卡在半路

### 一个参数时，确实会卡

先看最简单的情况：只有一个参数 w，损失随 w 变化的曲线有两个谷（图 4）。

<figure class="dd dd-figure">
<div class="dd-scroll">
<svg class="dd-svg" viewBox="0 0 800 340" xmlns="http://www.w3.org/2000/svg" role="img" aria-labelledby="gw3-title gw3-desc">
<title id="gw3-title">一个参数时，起点决定掉进哪个谷</title>
<desc id="gw3-desc">一条有两个谷的损失曲线。从 −0.6 出发的点一路走到左边更深的谷，损失 −5.64；从 0.6 出发的点走到右边较浅的谷，损失 −2.44，因为那里左右两边都往上，梯度下降停在那里出不来。</desc>
<defs>
<marker id="gw3-arrow" markerWidth="8" markerHeight="8" refX="7" refY="4" orient="auto"><path class="dd-head" d="M1 1 L7 4 L1 7"/></marker>
<marker id="gw3-arrow-acc" markerWidth="8" markerHeight="8" refX="7" refY="4" orient="auto"><path class="dd-head-acc" d="M1 1 L7 4 L1 7"/></marker>
</defs>
<rect class="dd-frame" width="800" height="340" rx="10"/>
<path class="dd-edge" style="stroke-width:2.2;fill:none" d="M84.7 172.7 L86.7 177.1 L88.8 181.4 L90.9 185.4 L93.0 189.3 L95.0 193.1 L97.1 196.7 L99.2 200.1 L101.2 203.4 L103.3 206.6 L105.4 209.6 L107.4 212.4 L109.5 215.1 L111.6 217.7 L113.7 220.2 L115.7 222.5 L117.8 224.7 L119.9 226.7 L121.9 228.7 L124.0 230.5 L126.1 232.2 L128.2 233.7 L130.2 235.2 L132.3 236.5 L134.4 237.8 L136.4 238.9 L138.5 239.9 L140.6 240.8 L142.7 241.6 L144.7 242.3 L146.8 242.9 L148.9 243.5 L150.9 243.9 L153.0 244.2 L155.1 244.5 L157.1 244.6 L159.2 244.7 L161.3 244.7 L163.4 244.6 L165.4 244.4 L167.5 244.2 L169.6 243.8 L171.6 243.4 L173.7 243.0 L175.8 242.4 L177.9 241.8 L179.9 241.2 L182.0 240.4 L184.1 239.6 L186.1 238.8 L188.2 237.9 L190.3 236.9 L192.4 235.9 L194.4 234.8 L196.5 233.7 L198.6 232.5 L200.6 231.3 L202.7 230.0 L204.8 228.7 L206.8 227.4 L208.9 226.0 L211.0 224.6 L213.1 223.1 L215.1 221.6 L217.2 220.0 L219.3 218.5 L221.3 216.9 L223.4 215.2 L225.5 213.6 L227.6 211.9 L229.6 210.2 L231.7 208.4 L233.8 206.7 L235.8 204.9 L237.9 203.1 L240.0 201.3 L242.1 199.5 L244.1 197.6 L246.2 195.8 L248.3 193.9 L250.3 192.0 L252.4 190.1 L254.5 188.2 L256.6 186.3 L258.6 184.4 L260.7 182.4 L262.8 180.5 L264.8 178.6 L266.9 176.7 L269.0 174.7 L271.0 172.8 L273.1 170.9 L275.2 169.0 L277.3 167.0 L279.3 165.1 L281.4 163.2 L283.5 161.3 L285.5 159.4 L287.6 157.5 L289.7 155.6 L291.8 153.8 L293.8 151.9 L295.9 150.1 L298.0 148.3 L300.0 146.5 L302.1 144.7 L304.2 142.9 L306.3 141.1 L308.3 139.4 L310.4 137.6 L312.5 135.9 L314.5 134.3 L316.6 132.6 L318.7 130.9 L320.7 129.3 L322.8 127.7 L324.9 126.1 L327.0 124.6 L329.0 123.1 L331.1 121.6 L333.2 120.1 L335.2 118.7 L337.3 117.2 L339.4 115.8 L341.5 114.5 L343.5 113.2 L345.6 111.8 L347.7 110.6 L349.7 109.3 L351.8 108.1 L353.9 106.9 L356.0 105.8 L358.0 104.7 L360.1 103.6 L362.2 102.5 L364.2 101.5 L366.3 100.5 L368.4 99.6 L370.5 98.6 L372.5 97.8 L374.6 96.9 L376.7 96.1 L378.7 95.3 L380.8 94.6 L382.9 93.9 L384.9 93.2 L387.0 92.5 L389.1 91.9 L391.2 91.4 L393.2 90.8 L395.3 90.3 L397.4 89.9 L399.4 89.4 L401.5 89.1 L403.6 88.7 L405.7 88.4 L407.7 88.1 L409.8 87.9 L411.9 87.6 L413.9 87.5 L416.0 87.3 L418.1 87.2 L420.2 87.2 L422.2 87.1 L424.3 87.1 L426.4 87.2 L428.4 87.2 L430.5 87.3 L432.6 87.5 L434.6 87.6 L436.7 87.9 L438.8 88.1 L440.9 88.4 L442.9 88.7 L445.0 89.0 L447.1 89.4 L449.1 89.8 L451.2 90.2 L453.3 90.7 L455.4 91.2 L457.4 91.7 L459.5 92.2 L461.6 92.8 L463.6 93.4 L465.7 94.1 L467.8 94.7 L469.9 95.4 L471.9 96.1 L474.0 96.9 L476.1 97.6 L478.1 98.4 L480.2 99.2 L482.3 100.1 L484.4 100.9 L486.4 101.8 L488.5 102.7 L490.6 103.6 L492.6 104.6 L494.7 105.5 L496.8 106.5 L498.8 107.5 L500.9 108.5 L503.0 109.5 L505.1 110.5 L507.1 111.6 L509.2 112.7 L511.3 113.7 L513.3 114.8 L515.4 115.9 L517.5 117.0 L519.6 118.1 L521.6 119.2 L523.7 120.4 L525.8 121.5 L527.8 122.6 L529.9 123.7 L532.0 124.9 L534.1 126.0 L536.1 127.1 L538.2 128.3 L540.3 129.4 L542.3 130.5 L544.4 131.6 L546.5 132.8 L548.5 133.9 L550.6 135.0 L552.7 136.0 L554.8 137.1 L556.8 138.2 L558.9 139.2 L561.0 140.2 L563.0 141.3 L565.1 142.2 L567.2 143.2 L569.3 144.2 L571.3 145.1 L573.4 146.0 L575.5 146.9 L577.5 147.7 L579.6 148.6 L581.7 149.3 L583.8 150.1 L585.8 150.8 L587.9 151.5 L590.0 152.2 L592.0 152.8 L594.1 153.4 L596.2 153.9 L598.3 154.4 L600.3 154.9 L602.4 155.3 L604.5 155.6 L606.5 155.9 L608.6 156.2 L610.7 156.4 L612.7 156.5 L614.8 156.6 L616.9 156.6 L619.0 156.6 L621.0 156.5 L623.1 156.3 L625.2 156.1 L627.2 155.8 L629.3 155.5 L631.4 155.0 L633.5 154.5 L635.5 153.9 L637.6 153.3 L639.7 152.5 L641.7 151.7 L643.8 150.8 L645.9 149.8 L648.0 148.7 L650.0 147.6 L652.1 146.3 L654.2 144.9 L656.2 143.5 L658.3 141.9 L660.4 140.3 L662.4 138.5 L664.5 136.7 L666.6 134.7 L668.7 132.6 L670.7 130.4 L672.8 128.1 L674.9 125.7 L676.9 123.2 L679.0 120.5 L681.1 117.7 L683.2 114.8 L685.2 111.8 L687.3 108.6 L689.4 105.3 L691.4 101.8 L693.5 98.3 L695.6 94.5 L697.7 90.7 L699.7 86.7 L701.8 82.5 L703.9 78.2"/>
<circle class="dd-dot" cx="331.2" cy="121.5" r="4.5"/>
<circle class="dd-dot" cx="314.1" cy="134.6" r="4.5"/>
<circle class="dd-dot" cx="294.7" cy="151.1" r="4.5"/>
<circle class="dd-dot" cx="273.5" cy="170.5" r="4.5"/>
<circle class="dd-dot" cx="251.3" cy="191.1" r="4.5"/>
<circle class="dd-dot" cx="229.5" cy="210.3" r="4.5"/>
<circle class="dd-dot" cx="209.7" cy="225.5" r="4.5"/>
<circle class="dd-dot" cx="193.2" cy="235.4" r="4.5"/>
<circle class="dd-dot" cx="180.9" cy="240.8" r="4.5"/>
<circle class="dd-dot" cx="167.2" cy="244.2" r="4.5"/>
<circle class="dd-dot" cx="162.2" cy="244.6" r="4.5"/>
<circle class="dd-dot" cx="160.6" cy="244.7" r="4.5"/>
<circle class="dd-dot" cx="160.2" cy="244.7" r="4.5"/>
<circle class="dd-dot" cx="160.0" cy="244.7" r="4.5"/>
<circle class="dd-dot" cx="160.0" cy="244.7" r="4.5"/>
<circle class="dd-dot" cx="160.0" cy="244.7" r="4.5"/>
<circle class="dd-dot" cx="160.0" cy="244.7" r="4.5"/>
<circle class="dd-dot" cx="160.0" cy="244.7" r="4.5"/>
<circle class="dd-dot" cx="160.0" cy="244.7" r="4.5"/>
<circle class="dd-fill-acc" cx="468.8" cy="95.1" r="4.5"/>
<circle class="dd-fill-acc" cx="476.7" cy="97.9" r="4.5"/>
<circle class="dd-fill-acc" cx="485.8" cy="101.5" r="4.5"/>
<circle class="dd-fill-acc" cx="495.9" cy="106.1" r="4.5"/>
<circle class="dd-fill-acc" cx="507.2" cy="111.6" r="4.5"/>
<circle class="dd-fill-acc" cx="519.4" cy="118.0" r="4.5"/>
<circle class="dd-fill-acc" cx="532.2" cy="125.0" r="4.5"/>
<circle class="dd-fill-acc" cx="545.2" cy="132.1" r="4.5"/>
<circle class="dd-fill-acc" cx="558.1" cy="138.8" r="4.5"/>
<circle class="dd-fill-acc" cx="580.8" cy="149.0" r="4.5"/>
<circle class="dd-fill-acc" cx="597.2" cy="154.2" r="4.5"/>
<circle class="dd-fill-acc" cx="607.1" cy="156.0" r="4.5"/>
<circle class="dd-fill-acc" cx="612.3" cy="156.5" r="4.5"/>
<circle class="dd-fill-acc" cx="614.8" cy="156.6" r="4.5"/>
<circle class="dd-fill-acc" cx="616.0" cy="156.6" r="4.5"/>
<circle class="dd-fill-acc" cx="616.5" cy="156.6" r="4.5"/>
<circle class="dd-fill-acc" cx="616.7" cy="156.6" r="4.5"/>
<circle class="dd-fill-acc" cx="616.8" cy="156.6" r="4.5"/>
<circle class="dd-fill-acc" cx="616.8" cy="156.6" r="4.5"/>
<text class="t-name" x="323.2" y="105.5" text-anchor="end">从 −0.6 出发</text>
<text class="t-name t-acc" x="476.8" y="79.1">从 0.6 出发</text>
<text class="t-note" x="160.3" y="268.7" text-anchor="middle">停在 −2.09，损失 −5.64（最低）</text>
<text class="t-note t-acc" x="616.7" y="180.6" text-anchor="middle">停在 1.89，损失 −2.44</text>
<text class="t-say" x="616.7" y="208.6" text-anchor="middle">两边都往上，梯度下降出不去</text>
<text class="t-note" x="56" y="324">只有一个参数 w，损失 = w⁴/4 − 2w² + 0.8w，学习率 0.05</text>
</svg>
</div>
<figcaption>图 4 · 只有一个参数时，“卡在一个不够好的谷里”是真的会发生的</figcaption>
</figure>

从 −0.6 出发，坡度一直指向左边，走到更深的谷，损失 −5.64。从 0.6 出发，坡度指向右边，走到右边那个浅一点的谷，损失 −2.44。到了那里，往左往右都是上坡，坡度是 0，梯度下降就停了。它并不知道左边还有一个更深的谷，因为它只看脚下。

这种四面都是上坡、但又不是最低的地方，叫**局部极小**。如果大模型的损失地形里到处都是这种坑，训练就会像人凭局部经验改方案一样，常常停在一个“改哪儿都更差、但其实不是最好”的方案上。

那真实的网络里到底卡不卡？我们直接试。

### 实验：神经元越多，越不容易卡

第一个任务是 XOR，也就是“异或”：平面上 4 个点，(−1, −1) 和 (1, 1) 标 0，(−1, 1) 和 (1, −1) 标 1。它很小，但一条直线分不开这两类，所以网络里至少要有 2 个神经元才学得会。第二个任务是上面那两条螺旋。两个任务都用全部样本一起算梯度，走法里没有任何随机性，唯一随机的是参数的起点。我们只改隐藏神经元的个数，每种个数从很多个随机起点各训练一遍，看有几次学会（表 1）。

<figure class="dd dd-figure dt-figure">
<div class="dt-panel">
<div class="dt-title" id="gwt1-title">同一个任务，只改隐藏神经元个数，随机起点各跑一遍</div>
<div class="dt-scroll">
<table class="dt" aria-labelledby="gwt1-title">
<colgroup><col style="width:22%"><col><col class="hl"><col></colgroup>
<thead><tr><th scope="col">神经元</th><th scope="col">参数个数</th><th scope="col" class="acc">学会的次数</th><th scope="col">最差一次停在</th></tr></thead>
<tbody>
<tr class="dt-group"><th colspan="4" scope="rowgroup">XOR（4 个点）</th></tr>
<tr><th scope="row">1</th><td class="mono num">5</td><td class="mono num bad">0 / 100</td><td class="mono num">0.693</td></tr>
<tr><th scope="row">2</th><td class="mono num">9</td><td class="mono num warn">74 / 100</td><td class="mono num">0.347</td></tr>
<tr><th scope="row">3</th><td class="mono num">13</td><td class="mono num warn">98 / 100</td><td class="mono num">0.347</td></tr>
<tr><th scope="row">4</th><td class="mono num">17</td><td class="mono num good">100 / 100</td><td class="mono num">0.001</td></tr>
<tr><th scope="row">8</th><td class="mono num">33</td><td class="mono num good">100 / 100</td><td class="mono num">0.001</td></tr>
<tr><th scope="row">32</th><td class="mono num">129</td><td class="mono num good">100 / 100</td><td class="mono num">0.000</td></tr>
<tr class="dt-group"><th colspan="4" scope="rowgroup">两条螺旋（200 个点）</th></tr>
<tr><th scope="row">2</th><td class="mono num">9</td><td class="mono num bad">0 / 40</td><td class="mono num">0.679</td></tr>
<tr><th scope="row">4</th><td class="mono num">17</td><td class="mono num bad">0 / 40</td><td class="mono num">0.340</td></tr>
<tr><th scope="row">8</th><td class="mono num">33</td><td class="mono num good">40 / 40</td><td class="mono num">0.004</td></tr>
<tr><th scope="row">16</th><td class="mono num">65</td><td class="mono num good">40 / 40</td><td class="mono num">0.001</td></tr>
<tr><th scope="row">32</th><td class="mono num">129</td><td class="mono num good">40 / 40</td><td class="mono num">0.001</td></tr>
<tr><th scope="row">64</th><td class="mono num">257</td><td class="mono num good">40 / 40</td><td class="mono num">0.001</td></tr>
</tbody>
</table>
</div>
<p class="dt-note">学会：训练后平均损失低于 0.05。XOR 每种 100 个起点、走 5000 步；螺旋每种 40 个起点、走 2 万步。学习率都是 0.5，全部样本一起算梯度，没有随机性。</p>
</div>
<figcaption>表 1 · 神经元少时经常卡住；多给几个，同样的走法每次都能走到底</figcaption>
</figure>

规律很清楚。XOR 用 2 个神经元时，100 个起点里有 26 个卡住了，每个都停在同一个损失 0.347；用 3 个时只剩 2 个卡住；4 个以上，100 个起点全部学会。螺旋也一样：2 个、4 个神经元一次都学不会，8 个以上次次学会。**同样的任务、同样的走法，只是参数多了，卡住就没了。**

这和直觉相反。参数多了，要调的东西更多，按说应该更难才对。为什么反而更容易？

### 卡住的地方长什么样

我们去看 XOR 那 26 个卡住的起点。它们停下的样子都一样：4 个点里 2 个判断对了，另外 2 个判断成“一半一半”，平均损失 0.347。再多走 10 万步，损失只从 0.3470 降到 0.3466，确实是停住了。

怎么知道四周有没有下坡？对一个参数，看曲线是向上弯还是向下弯就行。对 9 个参数，可以算出损失在各个方向上的弯曲程度：往某个方向挪一点，损失像开口向上的抛物线那样变大，就是向上弯；像开口向下的抛物线那样变小，就是向下弯。结果是：9 个方向里，1 个向上弯，8 个几乎完全平，**没有一个向下弯**。这是一片四周没有下坡的平地，梯度下降在这里没有任何方向可走。

现在给这个卡住的网络加第 3 个神经元，它的输入权重和输出权重全部设成 0。这个神经元什么也不输出，网络算出来的结果和原来一模一样，损失还是 0.3466。可是参数从 9 个变成了 13 个，我们再量一次各个方向的弯曲：**多出了一个向下弯的方向**（图 5）。

<figure class="dd dd-figure">
<div class="dd-scroll">
<svg class="dd-svg" viewBox="0 0 800 412" xmlns="http://www.w3.org/2000/svg" role="img" aria-labelledby="gw4-title gw4-desc">
<title id="gw4-title">卡住的点加一个神经元以后变成了鞍点</title>
<desc id="gw4-desc">左边是一张损失的热图，横轴是新加神经元的输入权重，纵轴是它的输出权重，范围都是 −3 到 3。中心是原来卡住的点，损失 0.3466。沿两条虚线（只改一个权重）颜色不变；两个权重一正一负一起变大时颜色变深，损失降到 0.025。</desc>
<defs>
<marker id="gw4-arrow" markerWidth="8" markerHeight="8" refX="7" refY="4" orient="auto"><path class="dd-head" d="M1 1 L7 4 L1 7"/></marker>
<marker id="gw4-arrow-acc" markerWidth="8" markerHeight="8" refX="7" refY="4" orient="auto"><path class="dd-head-acc" d="M1 1 L7 4 L1 7"/></marker>
</defs>
<rect class="dd-frame" width="800" height="412" rx="10"/>
<rect class="dd-zone" x="16" y="16" width="380" height="380" rx="8"/>
<text class="t-zone" x="32" y="44"></text>
<rect class="dd-fill-acc" x="236.0" y="344.0" width="12.3" height="12.3" fill-opacity="0.37"/>
<rect class="dd-fill-acc" x="248.0" y="344.0" width="12.3" height="12.3" fill-opacity="0.58"/>
<rect class="dd-fill-acc" x="260.0" y="344.0" width="12.3" height="12.3" fill-opacity="0.68"/>
<rect class="dd-fill-acc" x="272.0" y="344.0" width="12.3" height="12.3" fill-opacity="0.73"/>
<rect class="dd-fill-acc" x="284.0" y="344.0" width="12.3" height="12.3" fill-opacity="0.76"/>
<rect class="dd-fill-acc" x="296.0" y="344.0" width="12.3" height="12.3" fill-opacity="0.77"/>
<rect class="dd-fill-acc" x="308.0" y="344.0" width="12.3" height="12.3" fill-opacity="0.78"/>
<rect class="dd-fill-acc" x="320.0" y="344.0" width="12.3" height="12.3" fill-opacity="0.78"/>
<rect class="dd-fill-acc" x="332.0" y="344.0" width="12.3" height="12.3" fill-opacity="0.79"/>
<rect class="dd-fill-acc" x="344.0" y="344.0" width="12.3" height="12.3" fill-opacity="0.79"/>
<rect class="dd-fill-acc" x="356.0" y="344.0" width="12.3" height="12.3" fill-opacity="0.79"/>
<rect class="dd-fill-acc" x="368.0" y="344.0" width="12.3" height="12.3" fill-opacity="0.79"/>
<rect class="dd-fill-acc" x="236.0" y="332.0" width="12.3" height="12.3" fill-opacity="0.34"/>
<rect class="dd-fill-acc" x="248.0" y="332.0" width="12.3" height="12.3" fill-opacity="0.55"/>
<rect class="dd-fill-acc" x="260.0" y="332.0" width="12.3" height="12.3" fill-opacity="0.65"/>
<rect class="dd-fill-acc" x="272.0" y="332.0" width="12.3" height="12.3" fill-opacity="0.71"/>
<rect class="dd-fill-acc" x="284.0" y="332.0" width="12.3" height="12.3" fill-opacity="0.74"/>
<rect class="dd-fill-acc" x="296.0" y="332.0" width="12.3" height="12.3" fill-opacity="0.75"/>
<rect class="dd-fill-acc" x="308.0" y="332.0" width="12.3" height="12.3" fill-opacity="0.76"/>
<rect class="dd-fill-acc" x="320.0" y="332.0" width="12.3" height="12.3" fill-opacity="0.77"/>
<rect class="dd-fill-acc" x="332.0" y="332.0" width="12.3" height="12.3" fill-opacity="0.77"/>
<rect class="dd-fill-acc" x="344.0" y="332.0" width="12.3" height="12.3" fill-opacity="0.77"/>
<rect class="dd-fill-acc" x="356.0" y="332.0" width="12.3" height="12.3" fill-opacity="0.77"/>
<rect class="dd-fill-acc" x="368.0" y="332.0" width="12.3" height="12.3" fill-opacity="0.77"/>
<rect class="dd-fill-acc" x="236.0" y="320.0" width="12.3" height="12.3" fill-opacity="0.32"/>
<rect class="dd-fill-acc" x="248.0" y="320.0" width="12.3" height="12.3" fill-opacity="0.51"/>
<rect class="dd-fill-acc" x="260.0" y="320.0" width="12.3" height="12.3" fill-opacity="0.62"/>
<rect class="dd-fill-acc" x="272.0" y="320.0" width="12.3" height="12.3" fill-opacity="0.68"/>
<rect class="dd-fill-acc" x="284.0" y="320.0" width="12.3" height="12.3" fill-opacity="0.71"/>
<rect class="dd-fill-acc" x="296.0" y="320.0" width="12.3" height="12.3" fill-opacity="0.73"/>
<rect class="dd-fill-acc" x="308.0" y="320.0" width="12.3" height="12.3" fill-opacity="0.74"/>
<rect class="dd-fill-acc" x="320.0" y="320.0" width="12.3" height="12.3" fill-opacity="0.74"/>
<rect class="dd-fill-acc" x="332.0" y="320.0" width="12.3" height="12.3" fill-opacity="0.75"/>
<rect class="dd-fill-acc" x="344.0" y="320.0" width="12.3" height="12.3" fill-opacity="0.75"/>
<rect class="dd-fill-acc" x="356.0" y="320.0" width="12.3" height="12.3" fill-opacity="0.75"/>
<rect class="dd-fill-acc" x="368.0" y="320.0" width="12.3" height="12.3" fill-opacity="0.75"/>
<rect class="dd-fill-acc" x="236.0" y="308.0" width="12.3" height="12.3" fill-opacity="0.29"/>
<rect class="dd-fill-acc" x="248.0" y="308.0" width="12.3" height="12.3" fill-opacity="0.48"/>
<rect class="dd-fill-acc" x="260.0" y="308.0" width="12.3" height="12.3" fill-opacity="0.59"/>
<rect class="dd-fill-acc" x="272.0" y="308.0" width="12.3" height="12.3" fill-opacity="0.65"/>
<rect class="dd-fill-acc" x="284.0" y="308.0" width="12.3" height="12.3" fill-opacity="0.68"/>
<rect class="dd-fill-acc" x="296.0" y="308.0" width="12.3" height="12.3" fill-opacity="0.70"/>
<rect class="dd-fill-acc" x="308.0" y="308.0" width="12.3" height="12.3" fill-opacity="0.71"/>
<rect class="dd-fill-acc" x="320.0" y="308.0" width="12.3" height="12.3" fill-opacity="0.72"/>
<rect class="dd-fill-acc" x="332.0" y="308.0" width="12.3" height="12.3" fill-opacity="0.72"/>
<rect class="dd-fill-acc" x="344.0" y="308.0" width="12.3" height="12.3" fill-opacity="0.72"/>
<rect class="dd-fill-acc" x="356.0" y="308.0" width="12.3" height="12.3" fill-opacity="0.72"/>
<rect class="dd-fill-acc" x="368.0" y="308.0" width="12.3" height="12.3" fill-opacity="0.73"/>
<rect class="dd-fill-acc" x="236.0" y="296.0" width="12.3" height="12.3" fill-opacity="0.26"/>
<rect class="dd-fill-acc" x="248.0" y="296.0" width="12.3" height="12.3" fill-opacity="0.44"/>
<rect class="dd-fill-acc" x="260.0" y="296.0" width="12.3" height="12.3" fill-opacity="0.55"/>
<rect class="dd-fill-acc" x="272.0" y="296.0" width="12.3" height="12.3" fill-opacity="0.61"/>
<rect class="dd-fill-acc" x="284.0" y="296.0" width="12.3" height="12.3" fill-opacity="0.64"/>
<rect class="dd-fill-acc" x="296.0" y="296.0" width="12.3" height="12.3" fill-opacity="0.66"/>
<rect class="dd-fill-acc" x="308.0" y="296.0" width="12.3" height="12.3" fill-opacity="0.68"/>
<rect class="dd-fill-acc" x="320.0" y="296.0" width="12.3" height="12.3" fill-opacity="0.68"/>
<rect class="dd-fill-acc" x="332.0" y="296.0" width="12.3" height="12.3" fill-opacity="0.69"/>
<rect class="dd-fill-acc" x="344.0" y="296.0" width="12.3" height="12.3" fill-opacity="0.69"/>
<rect class="dd-fill-acc" x="356.0" y="296.0" width="12.3" height="12.3" fill-opacity="0.69"/>
<rect class="dd-fill-acc" x="368.0" y="296.0" width="12.3" height="12.3" fill-opacity="0.69"/>
<rect class="dd-fill-acc" x="236.0" y="284.0" width="12.3" height="12.3" fill-opacity="0.23"/>
<rect class="dd-fill-acc" x="248.0" y="284.0" width="12.3" height="12.3" fill-opacity="0.40"/>
<rect class="dd-fill-acc" x="260.0" y="284.0" width="12.3" height="12.3" fill-opacity="0.50"/>
<rect class="dd-fill-acc" x="272.0" y="284.0" width="12.3" height="12.3" fill-opacity="0.56"/>
<rect class="dd-fill-acc" x="284.0" y="284.0" width="12.3" height="12.3" fill-opacity="0.60"/>
<rect class="dd-fill-acc" x="296.0" y="284.0" width="12.3" height="12.3" fill-opacity="0.62"/>
<rect class="dd-fill-acc" x="308.0" y="284.0" width="12.3" height="12.3" fill-opacity="0.63"/>
<rect class="dd-fill-acc" x="320.0" y="284.0" width="12.3" height="12.3" fill-opacity="0.64"/>
<rect class="dd-fill-acc" x="332.0" y="284.0" width="12.3" height="12.3" fill-opacity="0.65"/>
<rect class="dd-fill-acc" x="344.0" y="284.0" width="12.3" height="12.3" fill-opacity="0.65"/>
<rect class="dd-fill-acc" x="356.0" y="284.0" width="12.3" height="12.3" fill-opacity="0.65"/>
<rect class="dd-fill-acc" x="368.0" y="284.0" width="12.3" height="12.3" fill-opacity="0.65"/>
<rect class="dd-fill-acc" x="236.0" y="272.0" width="12.3" height="12.3" fill-opacity="0.20"/>
<rect class="dd-fill-acc" x="248.0" y="272.0" width="12.3" height="12.3" fill-opacity="0.35"/>
<rect class="dd-fill-acc" x="260.0" y="272.0" width="12.3" height="12.3" fill-opacity="0.45"/>
<rect class="dd-fill-acc" x="272.0" y="272.0" width="12.3" height="12.3" fill-opacity="0.51"/>
<rect class="dd-fill-acc" x="284.0" y="272.0" width="12.3" height="12.3" fill-opacity="0.55"/>
<rect class="dd-fill-acc" x="296.0" y="272.0" width="12.3" height="12.3" fill-opacity="0.57"/>
<rect class="dd-fill-acc" x="308.0" y="272.0" width="12.3" height="12.3" fill-opacity="0.58"/>
<rect class="dd-fill-acc" x="320.0" y="272.0" width="12.3" height="12.3" fill-opacity="0.59"/>
<rect class="dd-fill-acc" x="332.0" y="272.0" width="12.3" height="12.3" fill-opacity="0.60"/>
<rect class="dd-fill-acc" x="344.0" y="272.0" width="12.3" height="12.3" fill-opacity="0.60"/>
<rect class="dd-fill-acc" x="356.0" y="272.0" width="12.3" height="12.3" fill-opacity="0.60"/>
<rect class="dd-fill-acc" x="368.0" y="272.0" width="12.3" height="12.3" fill-opacity="0.60"/>
<rect class="dd-fill-acc" x="236.0" y="260.0" width="12.3" height="12.3" fill-opacity="0.17"/>
<rect class="dd-fill-acc" x="248.0" y="260.0" width="12.3" height="12.3" fill-opacity="0.30"/>
<rect class="dd-fill-acc" x="260.0" y="260.0" width="12.3" height="12.3" fill-opacity="0.39"/>
<rect class="dd-fill-acc" x="272.0" y="260.0" width="12.3" height="12.3" fill-opacity="0.45"/>
<rect class="dd-fill-acc" x="284.0" y="260.0" width="12.3" height="12.3" fill-opacity="0.49"/>
<rect class="dd-fill-acc" x="296.0" y="260.0" width="12.3" height="12.3" fill-opacity="0.51"/>
<rect class="dd-fill-acc" x="308.0" y="260.0" width="12.3" height="12.3" fill-opacity="0.52"/>
<rect class="dd-fill-acc" x="320.0" y="260.0" width="12.3" height="12.3" fill-opacity="0.53"/>
<rect class="dd-fill-acc" x="332.0" y="260.0" width="12.3" height="12.3" fill-opacity="0.53"/>
<rect class="dd-fill-acc" x="344.0" y="260.0" width="12.3" height="12.3" fill-opacity="0.54"/>
<rect class="dd-fill-acc" x="356.0" y="260.0" width="12.3" height="12.3" fill-opacity="0.54"/>
<rect class="dd-fill-acc" x="368.0" y="260.0" width="12.3" height="12.3" fill-opacity="0.54"/>
<rect class="dd-fill-acc" x="236.0" y="248.0" width="12.3" height="12.3" fill-opacity="0.14"/>
<rect class="dd-fill-acc" x="248.0" y="248.0" width="12.3" height="12.3" fill-opacity="0.25"/>
<rect class="dd-fill-acc" x="260.0" y="248.0" width="12.3" height="12.3" fill-opacity="0.33"/>
<rect class="dd-fill-acc" x="272.0" y="248.0" width="12.3" height="12.3" fill-opacity="0.38"/>
<rect class="dd-fill-acc" x="284.0" y="248.0" width="12.3" height="12.3" fill-opacity="0.41"/>
<rect class="dd-fill-acc" x="296.0" y="248.0" width="12.3" height="12.3" fill-opacity="0.43"/>
<rect class="dd-fill-acc" x="308.0" y="248.0" width="12.3" height="12.3" fill-opacity="0.45"/>
<rect class="dd-fill-acc" x="320.0" y="248.0" width="12.3" height="12.3" fill-opacity="0.45"/>
<rect class="dd-fill-acc" x="332.0" y="248.0" width="12.3" height="12.3" fill-opacity="0.46"/>
<rect class="dd-fill-acc" x="344.0" y="248.0" width="12.3" height="12.3" fill-opacity="0.46"/>
<rect class="dd-fill-acc" x="356.0" y="248.0" width="12.3" height="12.3" fill-opacity="0.46"/>
<rect class="dd-fill-acc" x="368.0" y="248.0" width="12.3" height="12.3" fill-opacity="0.46"/>
<rect class="dd-fill-acc" x="236.0" y="236.0" width="12.3" height="12.3" fill-opacity="0.11"/>
<rect class="dd-fill-acc" x="248.0" y="236.0" width="12.3" height="12.3" fill-opacity="0.19"/>
<rect class="dd-fill-acc" x="260.0" y="236.0" width="12.3" height="12.3" fill-opacity="0.26"/>
<rect class="dd-fill-acc" x="272.0" y="236.0" width="12.3" height="12.3" fill-opacity="0.30"/>
<rect class="dd-fill-acc" x="284.0" y="236.0" width="12.3" height="12.3" fill-opacity="0.33"/>
<rect class="dd-fill-acc" x="296.0" y="236.0" width="12.3" height="12.3" fill-opacity="0.35"/>
<rect class="dd-fill-acc" x="308.0" y="236.0" width="12.3" height="12.3" fill-opacity="0.36"/>
<rect class="dd-fill-acc" x="320.0" y="236.0" width="12.3" height="12.3" fill-opacity="0.36"/>
<rect class="dd-fill-acc" x="332.0" y="236.0" width="12.3" height="12.3" fill-opacity="0.37"/>
<rect class="dd-fill-acc" x="344.0" y="236.0" width="12.3" height="12.3" fill-opacity="0.37"/>
<rect class="dd-fill-acc" x="356.0" y="236.0" width="12.3" height="12.3" fill-opacity="0.37"/>
<rect class="dd-fill-acc" x="368.0" y="236.0" width="12.3" height="12.3" fill-opacity="0.37"/>
<rect class="dd-fill-acc" x="236.0" y="224.0" width="12.3" height="12.3" fill-opacity="0.07"/>
<rect class="dd-fill-acc" x="248.0" y="224.0" width="12.3" height="12.3" fill-opacity="0.13"/>
<rect class="dd-fill-acc" x="260.0" y="224.0" width="12.3" height="12.3" fill-opacity="0.18"/>
<rect class="dd-fill-acc" x="272.0" y="224.0" width="12.3" height="12.3" fill-opacity="0.21"/>
<rect class="dd-fill-acc" x="284.0" y="224.0" width="12.3" height="12.3" fill-opacity="0.23"/>
<rect class="dd-fill-acc" x="296.0" y="224.0" width="12.3" height="12.3" fill-opacity="0.25"/>
<rect class="dd-fill-acc" x="308.0" y="224.0" width="12.3" height="12.3" fill-opacity="0.25"/>
<rect class="dd-fill-acc" x="320.0" y="224.0" width="12.3" height="12.3" fill-opacity="0.26"/>
<rect class="dd-fill-acc" x="332.0" y="224.0" width="12.3" height="12.3" fill-opacity="0.26"/>
<rect class="dd-fill-acc" x="344.0" y="224.0" width="12.3" height="12.3" fill-opacity="0.27"/>
<rect class="dd-fill-acc" x="356.0" y="224.0" width="12.3" height="12.3" fill-opacity="0.27"/>
<rect class="dd-fill-acc" x="368.0" y="224.0" width="12.3" height="12.3" fill-opacity="0.27"/>
<rect class="dd-fill-acc" x="236.0" y="212.0" width="12.3" height="12.3" fill-opacity="0.04"/>
<rect class="dd-fill-acc" x="248.0" y="212.0" width="12.3" height="12.3" fill-opacity="0.07"/>
<rect class="dd-fill-acc" x="260.0" y="212.0" width="12.3" height="12.3" fill-opacity="0.09"/>
<rect class="dd-fill-acc" x="272.0" y="212.0" width="12.3" height="12.3" fill-opacity="0.11"/>
<rect class="dd-fill-acc" x="284.0" y="212.0" width="12.3" height="12.3" fill-opacity="0.12"/>
<rect class="dd-fill-acc" x="296.0" y="212.0" width="12.3" height="12.3" fill-opacity="0.13"/>
<rect class="dd-fill-acc" x="308.0" y="212.0" width="12.3" height="12.3" fill-opacity="0.14"/>
<rect class="dd-fill-acc" x="320.0" y="212.0" width="12.3" height="12.3" fill-opacity="0.14"/>
<rect class="dd-fill-acc" x="332.0" y="212.0" width="12.3" height="12.3" fill-opacity="0.14"/>
<rect class="dd-fill-acc" x="344.0" y="212.0" width="12.3" height="12.3" fill-opacity="0.14"/>
<rect class="dd-fill-acc" x="356.0" y="212.0" width="12.3" height="12.3" fill-opacity="0.14"/>
<rect class="dd-fill-acc" x="368.0" y="212.0" width="12.3" height="12.3" fill-opacity="0.14"/>
<rect class="dd-fill-acc" x="80.0" y="188.0" width="12.3" height="12.3" fill-opacity="0.14"/>
<rect class="dd-fill-acc" x="92.0" y="188.0" width="12.3" height="12.3" fill-opacity="0.14"/>
<rect class="dd-fill-acc" x="104.0" y="188.0" width="12.3" height="12.3" fill-opacity="0.14"/>
<rect class="dd-fill-acc" x="116.0" y="188.0" width="12.3" height="12.3" fill-opacity="0.14"/>
<rect class="dd-fill-acc" x="128.0" y="188.0" width="12.3" height="12.3" fill-opacity="0.14"/>
<rect class="dd-fill-acc" x="140.0" y="188.0" width="12.3" height="12.3" fill-opacity="0.14"/>
<rect class="dd-fill-acc" x="152.0" y="188.0" width="12.3" height="12.3" fill-opacity="0.13"/>
<rect class="dd-fill-acc" x="164.0" y="188.0" width="12.3" height="12.3" fill-opacity="0.12"/>
<rect class="dd-fill-acc" x="176.0" y="188.0" width="12.3" height="12.3" fill-opacity="0.11"/>
<rect class="dd-fill-acc" x="188.0" y="188.0" width="12.3" height="12.3" fill-opacity="0.09"/>
<rect class="dd-fill-acc" x="200.0" y="188.0" width="12.3" height="12.3" fill-opacity="0.07"/>
<rect class="dd-fill-acc" x="212.0" y="188.0" width="12.3" height="12.3" fill-opacity="0.04"/>
<rect class="dd-fill-acc" x="80.0" y="176.0" width="12.3" height="12.3" fill-opacity="0.27"/>
<rect class="dd-fill-acc" x="92.0" y="176.0" width="12.3" height="12.3" fill-opacity="0.27"/>
<rect class="dd-fill-acc" x="104.0" y="176.0" width="12.3" height="12.3" fill-opacity="0.27"/>
<rect class="dd-fill-acc" x="116.0" y="176.0" width="12.3" height="12.3" fill-opacity="0.26"/>
<rect class="dd-fill-acc" x="128.0" y="176.0" width="12.3" height="12.3" fill-opacity="0.26"/>
<rect class="dd-fill-acc" x="140.0" y="176.0" width="12.3" height="12.3" fill-opacity="0.25"/>
<rect class="dd-fill-acc" x="152.0" y="176.0" width="12.3" height="12.3" fill-opacity="0.25"/>
<rect class="dd-fill-acc" x="164.0" y="176.0" width="12.3" height="12.3" fill-opacity="0.23"/>
<rect class="dd-fill-acc" x="176.0" y="176.0" width="12.3" height="12.3" fill-opacity="0.21"/>
<rect class="dd-fill-acc" x="188.0" y="176.0" width="12.3" height="12.3" fill-opacity="0.18"/>
<rect class="dd-fill-acc" x="200.0" y="176.0" width="12.3" height="12.3" fill-opacity="0.13"/>
<rect class="dd-fill-acc" x="212.0" y="176.0" width="12.3" height="12.3" fill-opacity="0.07"/>
<rect class="dd-fill-acc" x="80.0" y="164.0" width="12.3" height="12.3" fill-opacity="0.37"/>
<rect class="dd-fill-acc" x="92.0" y="164.0" width="12.3" height="12.3" fill-opacity="0.37"/>
<rect class="dd-fill-acc" x="104.0" y="164.0" width="12.3" height="12.3" fill-opacity="0.37"/>
<rect class="dd-fill-acc" x="116.0" y="164.0" width="12.3" height="12.3" fill-opacity="0.37"/>
<rect class="dd-fill-acc" x="128.0" y="164.0" width="12.3" height="12.3" fill-opacity="0.36"/>
<rect class="dd-fill-acc" x="140.0" y="164.0" width="12.3" height="12.3" fill-opacity="0.36"/>
<rect class="dd-fill-acc" x="152.0" y="164.0" width="12.3" height="12.3" fill-opacity="0.35"/>
<rect class="dd-fill-acc" x="164.0" y="164.0" width="12.3" height="12.3" fill-opacity="0.33"/>
<rect class="dd-fill-acc" x="176.0" y="164.0" width="12.3" height="12.3" fill-opacity="0.30"/>
<rect class="dd-fill-acc" x="188.0" y="164.0" width="12.3" height="12.3" fill-opacity="0.26"/>
<rect class="dd-fill-acc" x="200.0" y="164.0" width="12.3" height="12.3" fill-opacity="0.19"/>
<rect class="dd-fill-acc" x="212.0" y="164.0" width="12.3" height="12.3" fill-opacity="0.11"/>
<rect class="dd-fill-acc" x="80.0" y="152.0" width="12.3" height="12.3" fill-opacity="0.46"/>
<rect class="dd-fill-acc" x="92.0" y="152.0" width="12.3" height="12.3" fill-opacity="0.46"/>
<rect class="dd-fill-acc" x="104.0" y="152.0" width="12.3" height="12.3" fill-opacity="0.46"/>
<rect class="dd-fill-acc" x="116.0" y="152.0" width="12.3" height="12.3" fill-opacity="0.46"/>
<rect class="dd-fill-acc" x="128.0" y="152.0" width="12.3" height="12.3" fill-opacity="0.45"/>
<rect class="dd-fill-acc" x="140.0" y="152.0" width="12.3" height="12.3" fill-opacity="0.45"/>
<rect class="dd-fill-acc" x="152.0" y="152.0" width="12.3" height="12.3" fill-opacity="0.43"/>
<rect class="dd-fill-acc" x="164.0" y="152.0" width="12.3" height="12.3" fill-opacity="0.41"/>
<rect class="dd-fill-acc" x="176.0" y="152.0" width="12.3" height="12.3" fill-opacity="0.38"/>
<rect class="dd-fill-acc" x="188.0" y="152.0" width="12.3" height="12.3" fill-opacity="0.33"/>
<rect class="dd-fill-acc" x="200.0" y="152.0" width="12.3" height="12.3" fill-opacity="0.25"/>
<rect class="dd-fill-acc" x="212.0" y="152.0" width="12.3" height="12.3" fill-opacity="0.14"/>
<rect class="dd-fill-acc" x="80.0" y="140.0" width="12.3" height="12.3" fill-opacity="0.54"/>
<rect class="dd-fill-acc" x="92.0" y="140.0" width="12.3" height="12.3" fill-opacity="0.54"/>
<rect class="dd-fill-acc" x="104.0" y="140.0" width="12.3" height="12.3" fill-opacity="0.54"/>
<rect class="dd-fill-acc" x="116.0" y="140.0" width="12.3" height="12.3" fill-opacity="0.53"/>
<rect class="dd-fill-acc" x="128.0" y="140.0" width="12.3" height="12.3" fill-opacity="0.53"/>
<rect class="dd-fill-acc" x="140.0" y="140.0" width="12.3" height="12.3" fill-opacity="0.52"/>
<rect class="dd-fill-acc" x="152.0" y="140.0" width="12.3" height="12.3" fill-opacity="0.51"/>
<rect class="dd-fill-acc" x="164.0" y="140.0" width="12.3" height="12.3" fill-opacity="0.49"/>
<rect class="dd-fill-acc" x="176.0" y="140.0" width="12.3" height="12.3" fill-opacity="0.45"/>
<rect class="dd-fill-acc" x="188.0" y="140.0" width="12.3" height="12.3" fill-opacity="0.39"/>
<rect class="dd-fill-acc" x="200.0" y="140.0" width="12.3" height="12.3" fill-opacity="0.30"/>
<rect class="dd-fill-acc" x="212.0" y="140.0" width="12.3" height="12.3" fill-opacity="0.17"/>
<rect class="dd-fill-acc" x="80.0" y="128.0" width="12.3" height="12.3" fill-opacity="0.60"/>
<rect class="dd-fill-acc" x="92.0" y="128.0" width="12.3" height="12.3" fill-opacity="0.60"/>
<rect class="dd-fill-acc" x="104.0" y="128.0" width="12.3" height="12.3" fill-opacity="0.60"/>
<rect class="dd-fill-acc" x="116.0" y="128.0" width="12.3" height="12.3" fill-opacity="0.60"/>
<rect class="dd-fill-acc" x="128.0" y="128.0" width="12.3" height="12.3" fill-opacity="0.59"/>
<rect class="dd-fill-acc" x="140.0" y="128.0" width="12.3" height="12.3" fill-opacity="0.58"/>
<rect class="dd-fill-acc" x="152.0" y="128.0" width="12.3" height="12.3" fill-opacity="0.57"/>
<rect class="dd-fill-acc" x="164.0" y="128.0" width="12.3" height="12.3" fill-opacity="0.55"/>
<rect class="dd-fill-acc" x="176.0" y="128.0" width="12.3" height="12.3" fill-opacity="0.51"/>
<rect class="dd-fill-acc" x="188.0" y="128.0" width="12.3" height="12.3" fill-opacity="0.45"/>
<rect class="dd-fill-acc" x="200.0" y="128.0" width="12.3" height="12.3" fill-opacity="0.35"/>
<rect class="dd-fill-acc" x="212.0" y="128.0" width="12.3" height="12.3" fill-opacity="0.20"/>
<rect class="dd-fill-acc" x="80.0" y="116.0" width="12.3" height="12.3" fill-opacity="0.65"/>
<rect class="dd-fill-acc" x="92.0" y="116.0" width="12.3" height="12.3" fill-opacity="0.65"/>
<rect class="dd-fill-acc" x="104.0" y="116.0" width="12.3" height="12.3" fill-opacity="0.65"/>
<rect class="dd-fill-acc" x="116.0" y="116.0" width="12.3" height="12.3" fill-opacity="0.65"/>
<rect class="dd-fill-acc" x="128.0" y="116.0" width="12.3" height="12.3" fill-opacity="0.64"/>
<rect class="dd-fill-acc" x="140.0" y="116.0" width="12.3" height="12.3" fill-opacity="0.63"/>
<rect class="dd-fill-acc" x="152.0" y="116.0" width="12.3" height="12.3" fill-opacity="0.62"/>
<rect class="dd-fill-acc" x="164.0" y="116.0" width="12.3" height="12.3" fill-opacity="0.60"/>
<rect class="dd-fill-acc" x="176.0" y="116.0" width="12.3" height="12.3" fill-opacity="0.56"/>
<rect class="dd-fill-acc" x="188.0" y="116.0" width="12.3" height="12.3" fill-opacity="0.50"/>
<rect class="dd-fill-acc" x="200.0" y="116.0" width="12.3" height="12.3" fill-opacity="0.40"/>
<rect class="dd-fill-acc" x="212.0" y="116.0" width="12.3" height="12.3" fill-opacity="0.23"/>
<rect class="dd-fill-acc" x="80.0" y="104.0" width="12.3" height="12.3" fill-opacity="0.69"/>
<rect class="dd-fill-acc" x="92.0" y="104.0" width="12.3" height="12.3" fill-opacity="0.69"/>
<rect class="dd-fill-acc" x="104.0" y="104.0" width="12.3" height="12.3" fill-opacity="0.69"/>
<rect class="dd-fill-acc" x="116.0" y="104.0" width="12.3" height="12.3" fill-opacity="0.69"/>
<rect class="dd-fill-acc" x="128.0" y="104.0" width="12.3" height="12.3" fill-opacity="0.68"/>
<rect class="dd-fill-acc" x="140.0" y="104.0" width="12.3" height="12.3" fill-opacity="0.68"/>
<rect class="dd-fill-acc" x="152.0" y="104.0" width="12.3" height="12.3" fill-opacity="0.66"/>
<rect class="dd-fill-acc" x="164.0" y="104.0" width="12.3" height="12.3" fill-opacity="0.64"/>
<rect class="dd-fill-acc" x="176.0" y="104.0" width="12.3" height="12.3" fill-opacity="0.61"/>
<rect class="dd-fill-acc" x="188.0" y="104.0" width="12.3" height="12.3" fill-opacity="0.55"/>
<rect class="dd-fill-acc" x="200.0" y="104.0" width="12.3" height="12.3" fill-opacity="0.44"/>
<rect class="dd-fill-acc" x="212.0" y="104.0" width="12.3" height="12.3" fill-opacity="0.26"/>
<rect class="dd-fill-acc" x="80.0" y="92.0" width="12.3" height="12.3" fill-opacity="0.73"/>
<rect class="dd-fill-acc" x="92.0" y="92.0" width="12.3" height="12.3" fill-opacity="0.72"/>
<rect class="dd-fill-acc" x="104.0" y="92.0" width="12.3" height="12.3" fill-opacity="0.72"/>
<rect class="dd-fill-acc" x="116.0" y="92.0" width="12.3" height="12.3" fill-opacity="0.72"/>
<rect class="dd-fill-acc" x="128.0" y="92.0" width="12.3" height="12.3" fill-opacity="0.72"/>
<rect class="dd-fill-acc" x="140.0" y="92.0" width="12.3" height="12.3" fill-opacity="0.71"/>
<rect class="dd-fill-acc" x="152.0" y="92.0" width="12.3" height="12.3" fill-opacity="0.70"/>
<rect class="dd-fill-acc" x="164.0" y="92.0" width="12.3" height="12.3" fill-opacity="0.68"/>
<rect class="dd-fill-acc" x="176.0" y="92.0" width="12.3" height="12.3" fill-opacity="0.65"/>
<rect class="dd-fill-acc" x="188.0" y="92.0" width="12.3" height="12.3" fill-opacity="0.59"/>
<rect class="dd-fill-acc" x="200.0" y="92.0" width="12.3" height="12.3" fill-opacity="0.48"/>
<rect class="dd-fill-acc" x="212.0" y="92.0" width="12.3" height="12.3" fill-opacity="0.29"/>
<rect class="dd-fill-acc" x="80.0" y="80.0" width="12.3" height="12.3" fill-opacity="0.75"/>
<rect class="dd-fill-acc" x="92.0" y="80.0" width="12.3" height="12.3" fill-opacity="0.75"/>
<rect class="dd-fill-acc" x="104.0" y="80.0" width="12.3" height="12.3" fill-opacity="0.75"/>
<rect class="dd-fill-acc" x="116.0" y="80.0" width="12.3" height="12.3" fill-opacity="0.75"/>
<rect class="dd-fill-acc" x="128.0" y="80.0" width="12.3" height="12.3" fill-opacity="0.74"/>
<rect class="dd-fill-acc" x="140.0" y="80.0" width="12.3" height="12.3" fill-opacity="0.74"/>
<rect class="dd-fill-acc" x="152.0" y="80.0" width="12.3" height="12.3" fill-opacity="0.73"/>
<rect class="dd-fill-acc" x="164.0" y="80.0" width="12.3" height="12.3" fill-opacity="0.71"/>
<rect class="dd-fill-acc" x="176.0" y="80.0" width="12.3" height="12.3" fill-opacity="0.68"/>
<rect class="dd-fill-acc" x="188.0" y="80.0" width="12.3" height="12.3" fill-opacity="0.62"/>
<rect class="dd-fill-acc" x="200.0" y="80.0" width="12.3" height="12.3" fill-opacity="0.51"/>
<rect class="dd-fill-acc" x="212.0" y="80.0" width="12.3" height="12.3" fill-opacity="0.32"/>
<rect class="dd-fill-acc" x="80.0" y="68.0" width="12.3" height="12.3" fill-opacity="0.77"/>
<rect class="dd-fill-acc" x="92.0" y="68.0" width="12.3" height="12.3" fill-opacity="0.77"/>
<rect class="dd-fill-acc" x="104.0" y="68.0" width="12.3" height="12.3" fill-opacity="0.77"/>
<rect class="dd-fill-acc" x="116.0" y="68.0" width="12.3" height="12.3" fill-opacity="0.77"/>
<rect class="dd-fill-acc" x="128.0" y="68.0" width="12.3" height="12.3" fill-opacity="0.77"/>
<rect class="dd-fill-acc" x="140.0" y="68.0" width="12.3" height="12.3" fill-opacity="0.76"/>
<rect class="dd-fill-acc" x="152.0" y="68.0" width="12.3" height="12.3" fill-opacity="0.75"/>
<rect class="dd-fill-acc" x="164.0" y="68.0" width="12.3" height="12.3" fill-opacity="0.74"/>
<rect class="dd-fill-acc" x="176.0" y="68.0" width="12.3" height="12.3" fill-opacity="0.71"/>
<rect class="dd-fill-acc" x="188.0" y="68.0" width="12.3" height="12.3" fill-opacity="0.65"/>
<rect class="dd-fill-acc" x="200.0" y="68.0" width="12.3" height="12.3" fill-opacity="0.55"/>
<rect class="dd-fill-acc" x="212.0" y="68.0" width="12.3" height="12.3" fill-opacity="0.34"/>
<rect class="dd-fill-acc" x="80.0" y="56.0" width="12.3" height="12.3" fill-opacity="0.79"/>
<rect class="dd-fill-acc" x="92.0" y="56.0" width="12.3" height="12.3" fill-opacity="0.79"/>
<rect class="dd-fill-acc" x="104.0" y="56.0" width="12.3" height="12.3" fill-opacity="0.79"/>
<rect class="dd-fill-acc" x="116.0" y="56.0" width="12.3" height="12.3" fill-opacity="0.79"/>
<rect class="dd-fill-acc" x="128.0" y="56.0" width="12.3" height="12.3" fill-opacity="0.78"/>
<rect class="dd-fill-acc" x="140.0" y="56.0" width="12.3" height="12.3" fill-opacity="0.78"/>
<rect class="dd-fill-acc" x="152.0" y="56.0" width="12.3" height="12.3" fill-opacity="0.77"/>
<rect class="dd-fill-acc" x="164.0" y="56.0" width="12.3" height="12.3" fill-opacity="0.76"/>
<rect class="dd-fill-acc" x="176.0" y="56.0" width="12.3" height="12.3" fill-opacity="0.73"/>
<rect class="dd-fill-acc" x="188.0" y="56.0" width="12.3" height="12.3" fill-opacity="0.68"/>
<rect class="dd-fill-acc" x="200.0" y="56.0" width="12.3" height="12.3" fill-opacity="0.58"/>
<rect class="dd-fill-acc" x="212.0" y="56.0" width="12.3" height="12.3" fill-opacity="0.37"/>
<rect class="dd-edge" style="fill:none" x="80" y="56" width="300" height="300"/>
<line class="dd-edge dd-dash" x1="80" y1="230.0" x2="380" y2="230.0"/>
<line class="dd-edge dd-dash" x1="230.0" y1="56" x2="230.0" y2="356"/>
<circle class="dd-dot" cx="230.0" cy="230.0" r="6"/>
<text class="t-note" x="80" y="376" text-anchor="middle">−3</text>
<text class="t-note" x="380" y="376" text-anchor="middle">3</text>
<text class="t-note" x="230.0" y="376" text-anchor="middle">新神经元的输入权重</text>
<text class="t-note" x="70" y="230.0" text-anchor="middle" transform="rotate(-90 70 230.0)" dy="0">新神经元的输出权重</text>
<rect class="dd-zone" x="408" y="16" width="376" height="380" rx="8"/>
<text class="t-zone" x="424" y="44">卡住的点，加一个神经元之后</text>
<text class="t-name" x="424" y="72">中心圆点：原来卡住的地方</text>
<text class="t-note" x="424" y="94">第 3 个神经元的两个权重都是 0，网络算出来的</text>
<text class="t-note" x="424" y="116">和只有 2 个神经元时一模一样，损失 0.3466。</text>
<text class="t-name" x="424" y="160">虚线：只改其中一个权重</text>
<text class="t-note" x="424" y="182">沿着横线或竖线走到哪里，损失都还是 0.3466。</text>
<text class="t-note" x="424" y="204">13 个参数，单独改哪一个都降不下去。</text>
<text class="t-name t-acc" x="424" y="248">颜色越深：损失越低</text>
<text class="t-note t-acc" x="424" y="270">两个权重一正一负一起改，损失往下掉；</text>
<text class="t-note t-acc" x="424" y="292">改到 3 和 −3 时只剩 0.025。</text>
<text class="t-say" x="424" y="336">原来 9 个参数的时候，这里四周没有往下的方向；</text>
<text class="t-say" x="424" y="358">多了一个神经元，就多出一条下坡路。</text>
</svg>
</div>
<figcaption>图 5 · 单独改哪个参数都没用，两个一起改就有路：多出来的参数把谷底变成了鞍点</figcaption>
</figure>

图 5 里有个很像人改方案的细节。在这个点上，13 个参数单独改哪一个，从 −3 一直改到 3，损失都纹丝不动，最好还是 0.3466。原因是新神经元的两个权重互相“卡着”：输出权重是 0，输入权重怎么改都传不出去；输入权重是 0，这个神经元什么也没算出来，输出权重怎么改也没用。只有两个一起改，一正一负，损失才往下掉，一个改到 3、一个改到 −3 时只剩 0.025。

我们给 26 个卡住的网络都加上这样一个神经元，只是把它的权重设成很小的随机数（0.01 量级）而不是正好 0，然后接着训练。26 个全部走了出来：100 步后损失降到 0.03 到 0.08，2 万步后都在 0.0003 左右。小随机数的作用是让它不正好站在平地中央，一旦偏离一点点，向下弯的方向就会把这点偏离越放越大。真实训练里，随机初值和小批量带来的抖动，起的就是这个作用。

像这样“有的方向向上弯、有的方向向下弯”的点，叫**鞍点**，像马鞍：顺着马背方向是上坡，横着马背方向是下坡。梯度下降在鞍点附近会走得很慢，看起来像卡住了，但它不是死路。

### 为什么参数越多，死路越少

一个点要成为真正的死路，必须每个方向都向上弯，或者至少没有一个方向向下弯。方向越多，这个条件越难满足。

可以用一个粗糙的模型感受一下：假设每个点在各个方向上的弯曲是随机的，有的向上、有的向下。我们随机生成这样的弯曲，数一数“所有方向都向上弯”的比例：1 个方向时是 50%，2 个方向 14.7%，3 个方向 2.5%，4 个方向 0.26%，5 个方向 0.01%，到 6 个方向，20 万次里一次也没有出现。比例掉得比“每多一个方向减半”还快得多。

真实网络的弯曲当然不是随机的，所以这只是一个直觉。但研究者对大网络的分析得出了同样方向的结论：在损失还很高的地方，绝大多数停下来的点都是鞍点而不是死路[^dauphin]；网络越大，真正的局部极小越集中在损失很低、彼此差不多的一小段里[^choromanska]。换句话说，**参数多到一定程度以后，大部分“坑”都有出口，没有出口的坑本身就已经很深了。**

### 好的解不是一个，而是一大片

还有一个原因让训练不需要“找准”：好的解非常多。

XOR 用 8 个神经元时，100 个起点全部学会，但它们最后的参数值各不相同。原因之一是对称：把第 1 个神经元和第 3 个神经元的全部权重对调，网络算出来的东西完全不变；把一个神经元的输入权重和输出权重同时取反，结果也不变。8 个神经元的排列有 8! = 40320 种，每个还能取反，于是每一个解都至少有 40320 × 2⁸ ≈ 1032 万个参数不同、效果完全相同的“分身”。大模型每层有上万个神经元，这个数大到没有意义。

更进一步，研究者发现，从不同起点训出来的两个解，往往能找到一条连着它们的路径，路上的损失一直很低[^garipov][^draxler]。好的解不是散落在山里的几个针尖，而是连成一片的低洼地带。训练不需要找到“那一个”解，只需要走进这片洼地。

到这里，第一个疑问有了答案：训练能找到不错的解，不是因为它每一步都对，而是因为参数足够多时，死路很少、好解很多、连成一片，只要一直往下走，大概率会走进去。

**那么，走进去之后继续训练，能不能把每个参数都调到最优？**

## 三、继续训，能不能让每个参数都最优

### “每个参数都最优”指的是什么

先得想清楚这句话的意思。一个参数的最好取值，取决于其他所有参数现在是多少：图 5 里新神经元的输入权重，在输出权重为 0 时取什么都一样，在输出权重为 3 时，取 3 就比取 0 好得多。所以单独说“某个参数的最优值”没有意义，就像单独说“一个方案里预算的最优值”没有意义，它要看方案里的其他部分怎么定。

梯度下降最多能保证的是：走到坡度为 0 的地方，此时单独把任何一个参数往任何方向微调一点，损失都不会再降。但图 5 的那个点也满足这一条，13 个参数每一个单独看都“最优”，整体却离最好差得很远。所以**梯度下降从来没有保证过每个参数都最优**，它连“走到坡度为 0”都不保证。

### 大模型根本没走到底

真实的大模型训练，是在还没走到底的时候主动停下的。

训练一个模型的算力预算是固定的。2022 年 DeepMind 的 Chinchilla 研究算过：同样的预算，与其把一个大模型训到底，不如把模型做小一点、多喂一些数据[^chinchilla]。照这个算法，80 亿参数的模型大约喂 2000 亿个词元最划算。Meta 训练 Llama 3 时给 8B 和 70B 两个模型都喂到了 15 万亿个词元，远超这个数，并且说两个模型到这时性能仍在稳定提升[^llama3]。也就是说，它们发布时都还走在下坡路上。训练末期学习率会降到很小，那是为了让参数在当前位置停稳，而不是因为到了谷底。

### 训练损失越低越好吗

就算有无限的算力，能把训练损失一直降下去，也未必是好事。

我们换一个更像现实的实验：螺旋只给 120 个点，而且故意把其中 22 个点的标签标错，模拟真实数据里的噪声和错误。网络有 256 个神经元，1025 个参数，比点多得多。另外准备 4000 个训练时没见过、标签正确的点，用来检查模型是否真的学会了（图 6）。

<figure class="dd dd-figure">
<div class="dd-scroll">
<svg class="dd-svg" viewBox="0 0 800 344" xmlns="http://www.w3.org/2000/svg" role="img" aria-labelledby="gw5-title gw5-desc">
<title id="gw5-title">继续训练，训练损失一直降，新数据上的损失先降后升</title>
<desc id="gw5-desc">120 个训练点里有 22 个标签故意标错。训练损失从 0.71 一路降到 0.004；在 4000 个没见过、标签正确的点上，损失在第 3000 步最低（0.44），之后一路升到 2.97。</desc>
<defs>
<marker id="gw5-arrow" markerWidth="8" markerHeight="8" refX="7" refY="4" orient="auto"><path class="dd-head" d="M1 1 L7 4 L1 7"/></marker>
<marker id="gw5-arrow-acc" markerWidth="8" markerHeight="8" refX="7" refY="4" orient="auto"><path class="dd-head-acc" d="M1 1 L7 4 L1 7"/></marker>
</defs>
<rect class="dd-frame" width="800" height="344" rx="10"/>
<line class="dd-edge" x1="72" y1="296" x2="768" y2="296"/>
<line class="dd-edge" x1="72" y1="40" x2="72" y2="296"/>
<text class="t-note" x="62" y="300.0" text-anchor="end">0</text>
<text class="t-note" x="62" y="214.7" text-anchor="end">1</text>
<line class="dd-hair" x1="72" y1="210.7" x2="768" y2="210.7"/>
<text class="t-note" x="62" y="129.3" text-anchor="end">2</text>
<line class="dd-hair" x1="72" y1="125.3" x2="768" y2="125.3"/>
<text class="t-note" x="62" y="44.0" text-anchor="end">3</text>
<line class="dd-hair" x1="72" y1="40.0" x2="768" y2="40.0"/>
<text class="t-note" x="72.0" y="316" text-anchor="middle">500</text>
<text class="t-note" x="152.5" y="316" text-anchor="middle">1000</text>
<text class="t-note" x="280.1" y="316" text-anchor="middle">3000</text>
<text class="t-note" x="420.0" y="316" text-anchor="middle">1 万</text>
<text class="t-note" x="547.6" y="316" text-anchor="middle">3 万</text>
<text class="t-note" x="687.5" y="316" text-anchor="middle">10 万</text>
<text class="t-note" x="768.0" y="316" text-anchor="middle">20 万</text>
<text class="t-note" x="768" y="336" text-anchor="end">第几步（对数刻度）</text>
<text class="t-note" x="72" y="26">平均损失</text>
<path class="dd-edge" style="stroke-width:2;fill:none" d="M72.0 230.5 L152.5 255.8 L199.6 262.2 L233.0 266.1 L259.0 268.5 L280.1 270.3 L298.0 271.9 L313.6 273.3 L327.2 274.2 L339.5 275.6 L350.6 276.9 L360.7 277.9 L370.0 278.9 L378.6 279.7 L386.6 280.4 L394.1 281.1 L401.1 281.6 L407.8 282.2 L414.0 282.6 L420.0 283.0 L425.7 283.4 L431.1 283.8 L436.2 284.1 L441.2 284.4 L445.9 284.7 L450.5 285.0 L454.9 285.3 L459.1 285.5 L463.2 285.7 L467.1 286.0 L470.9 286.2 L474.6 286.4 L478.2 286.6 L481.6 286.8 L485.0 286.9 L488.3 287.1 L491.5 287.3 L494.6 287.4 L497.6 287.6 L500.5 287.7 L503.4 287.9 L506.2 288.0 L508.9 288.1 L511.6 288.3 L514.2 288.4 L516.8 288.5 L519.3 288.6 L521.7 288.7 L524.1 288.8 L526.4 288.9 L528.7 289.0 L531.0 289.1 L533.2 289.2 L535.4 289.3 L537.5 289.4 L539.6 289.5 L541.7 289.5 L543.7 289.6 L545.7 289.7 L547.6 289.8 L549.5 289.8 L551.4 289.9 L553.3 290.0 L555.1 290.0 L556.9 290.1 L558.7 290.2 L560.4 290.3 L562.2 290.3 L563.9 290.4 L565.5 290.5 L567.2 290.6 L568.8 290.7 L570.4 290.8 L572.0 290.9 L573.5 290.9 L575.1 291.0 L576.6 291.0 L578.1 291.1 L579.6 291.2 L581.0 291.2 L582.5 291.3 L583.9 291.3 L585.3 291.4 L586.7 291.4 L588.1 291.5 L589.4 291.5 L590.8 291.6 L592.1 291.6 L593.4 291.7 L594.7 291.7 L596.0 291.7 L597.3 291.8 L598.5 291.8 L599.8 291.9 L601.0 291.9 L602.2 292.0 L603.4 292.0 L604.6 292.0 L605.8 292.1 L607.0 292.1 L608.1 292.2 L609.3 292.2 L610.4 292.2 L611.5 292.3 L612.6 292.3 L613.7 292.3 L614.8 292.4 L615.9 292.4 L617.0 292.5 L618.0 292.5 L619.1 292.5 L620.1 292.6 L621.2 292.6 L622.2 292.6 L623.2 292.7 L624.2 292.7 L625.2 292.7 L626.2 292.8 L627.2 292.8 L628.1 292.8 L629.1 292.9 L630.1 292.9 L631.0 292.9 L631.9 292.9 L632.9 293.0 L633.8 293.0 L634.7 293.0 L635.6 293.1 L636.5 293.1 L637.4 293.1 L638.3 293.1 L639.2 293.2 L640.1 293.2 L641.0 293.2 L641.8 293.2 L642.7 293.3 L643.5 293.3 L644.4 293.3 L645.2 293.3 L646.0 293.4 L646.9 293.4 L647.7 293.4 L648.5 293.4 L649.3 293.5 L650.1 293.5 L650.9 293.5 L651.7 293.5 L652.5 293.5 L653.3 293.6 L654.1 293.6 L654.8 293.6 L655.6 293.6 L656.4 293.7 L657.1 293.7 L657.9 293.7 L658.6 293.7 L659.4 293.7 L660.1 293.8 L660.8 293.8 L661.6 293.8 L662.3 293.8 L663.0 293.8 L663.7 293.9 L664.4 293.9 L665.1 293.9 L665.8 293.9 L666.5 293.9 L667.2 294.0 L667.9 294.0 L668.6 294.0 L669.3 294.0 L670.0 294.0 L670.6 294.1 L671.3 294.1 L672.0 294.1 L672.6 294.1 L673.3 294.1 L673.9 294.1 L674.6 294.2 L675.2 294.2 L675.9 294.2 L676.5 294.2 L677.2 294.2 L677.8 294.3 L678.4 294.3 L679.1 294.3 L679.7 294.3 L680.3 294.3 L680.9 294.3 L681.5 294.4 L682.1 294.4 L682.7 294.4 L683.3 294.4 L683.9 294.4 L684.5 294.4 L685.1 294.5 L685.7 294.5 L686.3 294.5 L686.9 294.5 L687.5 294.5 L688.1 294.5 L688.6 294.6 L689.2 294.6 L689.8 294.6 L690.3 294.6 L690.9 294.6 L691.5 294.6 L692.0 294.7 L692.6 294.7 L693.1 294.7 L693.7 294.7 L694.2 294.7 L694.8 294.7 L695.3 294.7 L695.9 294.8 L696.4 294.8 L697.0 294.8 L697.5 294.8 L698.0 294.8 L698.6 294.8 L699.1 294.8 L699.6 294.8 L700.1 294.8 L700.6 294.9 L701.2 294.9 L701.7 294.9 L702.2 294.9 L702.7 294.9 L703.2 294.9 L703.7 294.9 L704.2 294.9 L704.7 294.9 L705.2 294.9 L705.7 294.9 L706.2 295.0 L706.7 295.0 L707.2 295.0 L707.7 295.0 L708.2 295.0 L708.7 295.0 L709.1 295.0 L709.6 295.0 L710.1 295.0 L710.6 295.0 L711.1 295.0 L711.5 295.1 L712.0 295.1 L712.5 295.1 L712.9 295.1 L713.4 295.1 L713.9 295.1 L714.3 295.1 L714.8 295.1 L715.2 295.1 L715.7 295.1 L716.2 295.1 L716.6 295.1 L717.1 295.1 L717.5 295.1 L718.0 295.2 L718.4 295.2 L718.8 295.2 L719.3 295.2 L719.7 295.2 L720.2 295.2 L720.6 295.2 L721.0 295.2 L721.5 295.2 L721.9 295.2 L722.3 295.2 L722.8 295.2 L723.2 295.2 L723.6 295.2 L724.1 295.2 L724.5 295.2 L724.9 295.3 L725.3 295.3 L725.7 295.3 L726.2 295.3 L726.6 295.3 L727.0 295.3 L727.4 295.3 L727.8 295.3 L728.2 295.3 L728.6 295.3 L729.0 295.3 L729.4 295.3 L729.8 295.3 L730.2 295.3 L730.6 295.3 L731.0 295.3 L731.4 295.3 L731.8 295.3 L732.2 295.3 L732.6 295.4 L733.0 295.4 L733.4 295.4 L733.8 295.4 L734.2 295.4 L734.6 295.4 L735.0 295.4 L735.4 295.4 L735.7 295.4 L736.1 295.4 L736.5 295.4 L736.9 295.4 L737.3 295.4 L737.6 295.4 L738.0 295.4 L738.4 295.4 L738.8 295.4 L739.1 295.4 L739.5 295.4 L739.9 295.4 L740.2 295.4 L740.6 295.4 L741.0 295.4 L741.4 295.4 L741.7 295.4 L742.1 295.5 L742.4 295.5 L742.8 295.5 L743.2 295.5 L743.5 295.5 L743.9 295.5 L744.2 295.5 L744.6 295.5 L744.9 295.5 L745.3 295.5 L745.7 295.5 L746.0 295.5 L746.4 295.5 L746.7 295.5 L747.1 295.5 L747.4 295.5 L747.7 295.5 L748.1 295.5 L748.4 295.5 L748.8 295.5 L749.1 295.5 L749.5 295.5 L749.8 295.5 L750.1 295.5 L750.5 295.5 L750.8 295.5 L751.2 295.5 L751.5 295.5 L751.8 295.5 L752.2 295.5 L752.5 295.5 L752.8 295.5 L753.2 295.6 L753.5 295.6 L753.8 295.6 L754.1 295.6 L754.5 295.6 L754.8 295.6 L755.1 295.6 L755.4 295.6 L755.8 295.6 L756.1 295.6 L756.4 295.6 L756.7 295.6 L757.0 295.6 L757.4 295.6 L757.7 295.6 L758.0 295.6 L758.3 295.6 L758.6 295.6 L758.9 295.6 L759.3 295.6 L759.6 295.6 L759.9 295.6 L760.2 295.6 L760.5 295.6 L760.8 295.6 L761.1 295.6 L761.4 295.6 L761.7 295.6 L762.0 295.6 L762.3 295.6 L762.7 295.6 L763.0 295.6 L763.3 295.6 L763.6 295.6 L763.9 295.6 L764.2 295.6 L764.5 295.6 L764.8 295.6 L765.1 295.6 L765.4 295.6 L765.7 295.6 L765.9 295.6 L766.2 295.6 L766.5 295.6 L766.8 295.6 L767.1 295.6 L767.4 295.6 L767.7 295.7 L768.0 295.7"/>
<path class="dd-edge-acc" style="stroke-width:2.4;fill:none" d="M72.0 224.2 L152.5 239.1 L199.6 252.6 L233.0 256.0 L259.0 257.6 L280.1 258.6 L298.0 257.6 L313.6 254.3 L327.2 254.0 L339.5 251.7 L350.6 249.3 L360.7 246.6 L370.0 244.1 L378.6 241.6 L386.6 239.0 L394.1 236.4 L401.1 234.0 L407.8 231.7 L414.0 229.4 L420.0 227.1 L425.7 224.8 L431.1 222.5 L436.2 220.2 L441.2 217.9 L445.9 215.7 L450.5 213.5 L454.9 211.5 L459.1 209.6 L463.2 207.8 L467.1 206.1 L470.9 204.5 L474.6 202.9 L478.2 201.4 L481.6 199.9 L485.0 198.4 L488.3 196.9 L491.5 195.5 L494.6 194.1 L497.6 192.7 L500.5 191.3 L503.4 190.0 L506.2 188.6 L508.9 187.3 L511.6 186.0 L514.2 184.7 L516.8 183.5 L519.3 182.2 L521.7 181.0 L524.1 179.8 L526.4 178.6 L528.7 177.5 L531.0 176.3 L533.2 175.2 L535.4 174.1 L537.5 173.1 L539.6 172.0 L541.7 171.0 L543.7 169.9 L545.7 168.9 L547.6 167.9 L549.5 166.8 L551.4 165.8 L553.3 164.8 L555.1 163.8 L556.9 162.8 L558.7 161.8 L560.4 160.7 L562.2 159.6 L563.9 158.6 L565.5 157.5 L567.2 156.5 L568.8 155.5 L570.4 154.5 L572.0 153.5 L573.5 152.6 L575.1 151.6 L576.6 150.7 L578.1 149.8 L579.6 149.0 L581.0 148.1 L582.5 147.3 L583.9 146.5 L585.3 145.7 L586.7 145.0 L588.1 144.3 L589.4 143.6 L590.8 142.9 L592.1 142.2 L593.4 141.6 L594.7 140.9 L596.0 140.3 L597.3 139.7 L598.5 139.1 L599.8 138.5 L601.0 138.0 L602.2 137.4 L603.4 136.9 L604.6 136.4 L605.8 135.9 L607.0 135.4 L608.1 134.9 L609.3 134.4 L610.4 133.9 L611.5 133.5 L612.6 133.1 L613.7 132.6 L614.8 132.2 L615.9 131.8 L617.0 131.4 L618.0 131.0 L619.1 130.7 L620.1 130.3 L621.2 129.9 L622.2 129.5 L623.2 129.2 L624.2 128.8 L625.2 128.4 L626.2 128.0 L627.2 127.7 L628.1 127.3 L629.1 126.9 L630.1 126.5 L631.0 126.1 L631.9 125.7 L632.9 125.4 L633.8 125.0 L634.7 124.6 L635.6 124.2 L636.5 123.8 L637.4 123.4 L638.3 123.0 L639.2 122.6 L640.1 122.2 L641.0 121.9 L641.8 121.5 L642.7 121.1 L643.5 120.7 L644.4 120.3 L645.2 119.8 L646.0 119.4 L646.9 119.0 L647.7 118.6 L648.5 118.1 L649.3 117.7 L650.1 117.2 L650.9 116.8 L651.7 116.3 L652.5 115.8 L653.3 115.4 L654.1 114.9 L654.8 114.4 L655.6 114.0 L656.4 113.5 L657.1 113.0 L657.9 112.6 L658.6 112.1 L659.4 111.6 L660.1 111.2 L660.8 110.7 L661.6 110.2 L662.3 109.7 L663.0 109.2 L663.7 108.7 L664.4 108.2 L665.1 107.7 L665.8 107.2 L666.5 106.6 L667.2 106.1 L667.9 105.6 L668.6 105.0 L669.3 104.5 L670.0 103.9 L670.6 103.3 L671.3 102.8 L672.0 102.2 L672.6 101.6 L673.3 101.1 L673.9 100.5 L674.6 99.9 L675.2 99.4 L675.9 98.8 L676.5 98.2 L677.2 97.7 L677.8 97.1 L678.4 96.5 L679.1 96.0 L679.7 95.5 L680.3 94.9 L680.9 94.4 L681.5 93.9 L682.1 93.5 L682.7 93.0 L683.3 92.5 L683.9 92.1 L684.5 91.7 L685.1 91.3 L685.7 90.9 L686.3 90.6 L686.9 90.3 L687.5 89.9 L688.1 89.6 L688.6 89.4 L689.2 89.1 L689.8 88.8 L690.3 88.6 L690.9 88.3 L691.5 88.1 L692.0 87.8 L692.6 87.6 L693.1 87.4 L693.7 87.2 L694.2 87.1 L694.8 87.0 L695.3 86.9 L695.9 86.8 L696.4 86.7 L697.0 86.6 L697.5 86.5 L698.0 86.4 L698.6 86.3 L699.1 86.2 L699.6 86.2 L700.1 86.1 L700.6 86.0 L701.2 85.9 L701.7 85.8 L702.2 85.8 L702.7 85.7 L703.2 85.6 L703.7 85.5 L704.2 85.4 L704.7 85.3 L705.2 85.2 L705.7 85.1 L706.2 84.9 L706.7 84.8 L707.2 84.7 L707.7 84.5 L708.2 84.4 L708.7 84.3 L709.1 84.1 L709.6 83.9 L710.1 83.8 L710.6 83.6 L711.1 83.4 L711.5 83.3 L712.0 83.1 L712.5 82.9 L712.9 82.7 L713.4 82.5 L713.9 82.3 L714.3 82.0 L714.8 81.8 L715.2 81.6 L715.7 81.4 L716.2 81.1 L716.6 80.9 L717.1 80.6 L717.5 80.4 L718.0 80.1 L718.4 79.9 L718.8 79.6 L719.3 79.4 L719.7 79.1 L720.2 78.9 L720.6 78.6 L721.0 78.4 L721.5 78.1 L721.9 77.9 L722.3 77.6 L722.8 77.4 L723.2 77.2 L723.6 76.9 L724.1 76.7 L724.5 76.4 L724.9 76.2 L725.3 75.9 L725.7 75.7 L726.2 75.4 L726.6 75.1 L727.0 74.9 L727.4 74.6 L727.8 74.3 L728.2 74.1 L728.6 73.8 L729.0 73.5 L729.4 73.2 L729.8 72.9 L730.2 72.7 L730.6 72.4 L731.0 72.1 L731.4 71.8 L731.8 71.5 L732.2 71.2 L732.6 70.9 L733.0 70.6 L733.4 70.3 L733.8 70.0 L734.2 69.8 L734.6 69.5 L735.0 69.2 L735.4 68.9 L735.7 68.6 L736.1 68.3 L736.5 68.0 L736.9 67.7 L737.3 67.4 L737.6 67.1 L738.0 66.8 L738.4 66.5 L738.8 66.3 L739.1 66.0 L739.5 65.7 L739.9 65.4 L740.2 65.1 L740.6 64.8 L741.0 64.5 L741.4 64.3 L741.7 64.0 L742.1 63.7 L742.4 63.4 L742.8 63.1 L743.2 62.9 L743.5 62.6 L743.9 62.3 L744.2 62.0 L744.6 61.7 L744.9 61.5 L745.3 61.2 L745.7 60.9 L746.0 60.6 L746.4 60.4 L746.7 60.1 L747.1 59.8 L747.4 59.5 L747.7 59.2 L748.1 59.0 L748.4 58.7 L748.8 58.4 L749.1 58.1 L749.5 57.9 L749.8 57.6 L750.1 57.3 L750.5 57.0 L750.8 56.8 L751.2 56.5 L751.5 56.2 L751.8 55.9 L752.2 55.7 L752.5 55.4 L752.8 55.1 L753.2 54.8 L753.5 54.6 L753.8 54.3 L754.1 54.0 L754.5 53.7 L754.8 53.5 L755.1 53.2 L755.4 52.9 L755.8 52.7 L756.1 52.4 L756.4 52.1 L756.7 51.9 L757.0 51.6 L757.4 51.3 L757.7 51.1 L758.0 50.8 L758.3 50.6 L758.6 50.3 L758.9 50.0 L759.3 49.8 L759.6 49.5 L759.9 49.3 L760.2 49.0 L760.5 48.7 L760.8 48.5 L761.1 48.2 L761.4 48.0 L761.7 47.7 L762.0 47.5 L762.3 47.2 L762.7 47.0 L763.0 46.7 L763.3 46.5 L763.6 46.2 L763.9 46.0 L764.2 45.7 L764.5 45.5 L764.8 45.3 L765.1 45.0 L765.4 44.8 L765.7 44.5 L765.9 44.3 L766.2 44.0 L766.5 43.8 L766.8 43.6 L767.1 43.3 L767.4 43.1 L767.7 42.9 L768.0 42.6"/>
<line class="dd-edge dd-dash" x1="280.1" y1="40" x2="280.1" y2="296"/>
<circle class="dd-fill-acc" cx="280.1" cy="258.6" r="5"/>
<text class="t-note" x="288.1" y="54">第 3000 步：新数据上最好，0.44</text>
<text class="t-name t-acc" x="628.1" y="65.6" text-anchor="end">没见过的 4000 个点：2.97</text>
<text class="t-name" x="764.0" y="283.7" text-anchor="end">训练用的 120 个点：0.004</text>
<text class="t-say" x="301.3" y="133.9">训练损失一路降到接近 0，</text>
<text class="t-say" x="301.3" y="153.9">连 22 个标错的点也背了下来。</text>
</svg>
</div>
<figcaption>图 6 · 评审标准写死以后，把它做到极致不等于把事情做好</figcaption>
</figure>

训练损失一路从 0.71 降到 0.004，120 个训练点全部“判断正确”，连那 22 个标错的点也照着错的标签背了下来。可是在没见过的点上，损失在第 3000 步最低，只有 0.44，之后一路升到 2.97；判断的正确率也从 77% 降到 73%。训练越久，模型越是在迁就那 22 个错误，越不像真正的螺旋。参数足够多的网络连完全随机的标签都能背下来[^zhang]，所以“训练损失降得越低”和“模型越好”并不是一回事。

这就是人类评审里常说的“为了过评审而过评审”。**评审标准一旦写死，把它做到极致，就不等于把事情做好。** 训练损失只是我们关心的东西的一个代理。所以实际训练会单独留一份不参与训练的数据，专门看模型在没见过的题上表现如何，在那里变差之前停下，这叫**早停**。大模型的数据多到每段文字大多只看一遍，背题的问题轻得多，但同样会留出验证数据一直盯着。

**所以训练里那一个写死的评审并不完美。人的那些来回评审，是被它取代了吗？**

## 四、人的评审去了哪里

没有被取代，而是被挪到了训练外面（图 7）。

<figure class="dd dd-figure">
<div class="dd-scroll">
<svg class="dd-svg" viewBox="0 0 800 400" xmlns="http://www.w3.org/2000/svg" role="img" aria-labelledby="gw6-title gw6-desc">
<title id="gw6-title">训练里外两层评审循环</title>
<desc id="gw6-desc">外层是人主导的循环：定评审标准（损失和数据）、定训练设置、训练、评测、人看回答、修改，再来一轮。中间虚线框是训练本身，用损失自动评审，每轮几秒，转几万到几十万轮。</desc>
<defs>
<marker id="gw6-arrow" markerWidth="8" markerHeight="8" refX="7" refY="4" orient="auto"><path class="dd-head" d="M1 1 L7 4 L1 7"/></marker>
<marker id="gw6-arrow-acc" markerWidth="8" markerHeight="8" refX="7" refY="4" orient="auto"><path class="dd-head-acc" d="M1 1 L7 4 L1 7"/></marker>
</defs>
<rect class="dd-frame" width="800" height="400" rx="10"/>
<rect class="dd-zone" x="16" y="16" width="768" height="368" rx="8"/>
<text class="t-zone" x="32" y="44">训练外面：人来评审　一轮几天到几个月，一共几十轮</text>
<rect class="dd-derived" x="248" y="64" width="304" height="200" rx="8"/>
<text class="t-zone t-acc" x="264" y="92">训练里面：损失来评审</text>
<rect class="dd-card" x="264" y="108" width="272" height="56" rx="6"/>
<text class="t-name t-acc" x="278" y="130">答题 → 损失 → 梯度 → 更新</text>
<text class="t-note t-acc" x="278" y="150">机器自动算，一共几万到几十万轮</text>
<text class="t-say" x="264" y="196">只认写好的那个损失，</text>
<text class="t-say" x="264" y="216">不会问这个损失定得对不对。</text>
<text class="t-note" x="264" y="250">虚线框：只在训练时转，训练完就停</text>
<rect class="dd-card" x="32" y="64" width="192" height="56" rx="6"/>
<text class="t-name" x="46" y="86">定评审标准</text>
<text class="t-note" x="46" y="106">写损失、选数据</text>
<rect class="dd-card" x="32" y="208" width="192" height="56" rx="6"/>
<text class="t-name" x="46" y="230">定怎么训</text>
<text class="t-note" x="46" y="250">多大、学习率、训多久</text>
<rect class="dd-card" x="576" y="64" width="192" height="56" rx="6"/>
<text class="t-name" x="590" y="86">评测</text>
<text class="t-note" x="590" y="106">没见过的题、基准测试</text>
<rect class="dd-card" x="576" y="208" width="192" height="56" rx="6"/>
<text class="t-name" x="590" y="230">人看回答</text>
<text class="t-note" x="590" y="250">排序、挑毛病</text>
<rect class="dd-card" x="248" y="296" width="304" height="56" rx="6"/>
<text class="t-name" x="262" y="318">修改</text>
<text class="t-note" x="262" y="338">改数据、改设置，或把人的排序训成新的评审</text>
<line class="dd-edge" x1="224" y1="92" x2="246" y2="92" marker-end="url(#gw6-arrow)"/>
<line class="dd-edge" x1="224" y1="236" x2="246" y2="236" marker-end="url(#gw6-arrow)"/>
<line class="dd-edge" x1="552" y1="92" x2="574" y2="92" marker-end="url(#gw6-arrow)"/>
<line class="dd-edge" x1="672" y1="120" x2="672" y2="206" marker-end="url(#gw6-arrow)"/>
<path class="dd-edge" style="fill:none" d="M672 264 V316 Q672 324 664 324 H554" marker-end="url(#gw6-arrow)"/>
<path class="dd-edge" style="fill:none" d="M248 324 H136 Q128 324 128 316 V266" marker-end="url(#gw6-arrow)"/>
<line class="dd-edge" x1="128" y1="208" x2="128" y2="122" marker-end="url(#gw6-arrow)"/>
</svg>
</div>
<figcaption>图 7 · 来回评审没有消失，它分成了两层：里层交给损失，外层还是人</figcaption>
</figure>

**定评审标准，本身就是最难的一轮评审。** 训练只会把写好的损失做低，不会问这个损失定得对不对。人做方案时“要考虑这么多”的那些因素，好用、安全、不胡说、符合语气，要么被折进损失和数据里，要么就不会被训练关心。语言模型的损失只是“下一个词猜得准不准”，训练语料选什么、怎么配比，决定了“猜得准”最后意味着什么。

**定怎么训，也只能一轮轮地试。** 模型多大、学习率多少、训多久，这些没有梯度可算。上一篇讲过，大模型团队通常先在小模型上试好这些选择，再放大；放大后训一次就要几周到几个月。

**训练完，还是要人来看。** 在没见过的题和基准测试上评测，让人读模型的回答、挑毛病，发现问题就改数据、改设置，重新训练。这一圈才是真正对应人改方案的那个循环，一轮很贵，一共也就转几十轮。

**人的评审也可以被做成损失。** ChatGPT 这一类模型的做法是：让人对同一个问题的几个回答排序，用这些排序训练一个专门打分的模型，叫奖励模型，再拿它当评审去训练语言模型[^instructgpt]。人的判断就这样被折进了一个可以求导的分数里，内层循环又能自动转起来了。可它同样会被钻空子：对着奖励模型一直优化，奖励模型给的分一直涨，换更可靠的方式去评，真实质量却是先升后降[^gao]，和图 6 是同一个形状。所以这个评审本身，还需要人在外层继续评审。

## 五、回到开头

现在可以回答开头那个问题了：这里面没有魔法，那些来回评审一轮也没有省掉，只是被拆开重新分配了。

能写成公式的那部分评审，交给了损失。它每一轮给出的不是“好或不好”，而是每个参数各自该往哪边改、改多少；一轮又便宜到可以转上几万几十万次，所以单独一轮改坏了没关系，下一轮会纠回来。参数多到一定程度以后，损失地形里的死路很少，好解很多、连成一片，只要一直往下走，大概率会走进好解的洼地。

但它找到的不是最优解，也不需要是。“每个参数都最优”本来就说不清；大模型都是在还在下坡时就主动停下；训练损失做到极致，反而可能离我们真正想要的更远。

写不成公式的那部分评审，评审标准定得对不对，模型训出来到底好不好，依然要人一轮一轮地看、一轮一轮地改，和我们平时改方案没什么两样。

## 小结

1. **训练里谁在评审？** 损失。它是一个写死的分数，每一步都评一次；反向传播再把它拆成每个参数的修改意见，精确到方向和大小。GPT-3 被这样评审了约 9.4 万轮。
2. **每一步都会变好吗？** 不会。每步只看一小批题，我们的实验里前 1000 步有近一半让整体变差，但前后平均一直往下：单轮改坏了，下一轮会纠回来。
3. **会不会卡在半路？** 参数少时会：XOR 用 2 个神经元，100 个起点卡住 26 个。参数多了，卡住的点会变成鞍点，多出的参数提供了下坡方向，4 个神经元以上就再没卡过。方向越多，“所有方向都向上弯”越难满足，所以大网络里死路少。
4. **为什么不需要找准？** 好解非常多。仅神经元的对调和取反，就让每个解有海量完全等价的分身；不同起点训出的解之间往往有一条低损失的路连着。
5. **继续训能让每个参数都最优吗？** 不能，也不必。一个参数的最好取值取决于其他参数；坡度为 0 只说明单独微调哪个参数都没用，鞍点也满足这一条。大模型在还在下坡时就主动停下；训练损失做到极致还可能背下数据里的错误，所以要用没见过的数据盯着、及时停下。
6. **人的评审去哪了？** 挪到了训练外面：定损失和数据、定训练设置、评测、人看回答、改了再训。人的判断还可以训成奖励模型，变成新的损失，但它同样会被钻空子，仍需人在外层把关。

## 参考资料
