机器学习 - 介绍
本页标题、字幕、摘要和阅读笔记为已保存的 AI 译文，可能有误。时间戳与来源一致，可展开查看原始字幕，并对照原视频核对。
Nando de Freitas 的第一堂机器学习课：为什么程序必须从数据中学习，并通过视觉、推荐、网络规模文本和动物感知来展示。
转录来源：Vidleaf上的自动语音识别（可能包含错误）。AI标点由Vidleaf添加；保留原始措辞和时间戳锚点。原始转录可在下方获取。
AI生成的摘要
Nando de Freitas 通过对比编程规则与从示例中学习的系统来介绍机器学习。一个变化盲视演示引出了视觉的困难，随后是涉及预测、推荐、语言和语音的例子。讲座强调数据的作用，并解释当缺乏明确专业知识或任务不断变化时，学习方法为何变得有用。后面的部分将神经系统、表示和深度学习背后的思想联系起来。该讲座是对该领域问题和直觉的介绍，其演示和历史例子保留在链接的章节中。
要点
- [8:26] 从示例中学习提供了一种不同于手工指定每条视觉规则的方法。
- [21:10] 当专家规则不可用或任务随时间变化时，学习方法很有用。
- [37:40] 讲座将神经网络和学习到的表示与识别任务的进展联系起来。
根据文字记录生成。请对照原始视频核查所链接的段落。简要说明和要点由 Vidleaf 于 2026-10-21 审核。
完整概述
本讲座以一个变化盲视演示开场，大多数观众没有注意到变化，重放视频仍然只揭示部分变化；演讲者认为眼睛每秒移动约三次，我们看到的大部分内容是想象出来的，这使得编程机器去观看极其困难。它介绍了机器学习、大数据革命、计算能力以及受生物系统启发的神经系统，对比了脆弱的手工编码视觉与使用大型图像集合和随机森林进行人脸和行人识别的基于示例的学习。一个“tufa”练习表明人们能从少量示例中快速学习新概念，具有强烈但有些主观的一致性，而机器无法做出这些归纳跳跃，而是提取特征、模式或统计量来进行预测。讨论的应用包括预测能源需求和销售、电影推荐、检测 HIV 突变、信用评分、医学诊断、排序和 Google 图像搜索（其改进来自人类点击），以及 Zite。机器学习被呈现为真正的 AI，在缺乏专业知识时需要它，例如火星机器人，以及对于不断变化的问题如网络推荐。该片段还认为规模使自动系统成为必要，引用数据集从 1960 的一百万词增长到 2006 的一万亿词，并将语言技术、Google 翻译和语音识别作为成功案例，其中数据是关键商品，也是真正的绊脚石。它对比了 Google 的倒排索引搜索和流行度排名与硬引用匹配，指出网络包含真相但也包含垃圾，最后指出学习还涉及干预和因果关系，深度学习推动了物体识别和语音识别的最新进展。最后一部分从联想记忆转向蜘蛛认知和位置细胞、V1 神经元和分层特征学习，包括 Hubel 和 Wiesel 的猫实验、神经元激活和输入重建，以及语音识别流程，下一节课将开始使用线性模型进行监督学习。
带时间戳的章节摘要
变化盲视与机器视觉的挑战
- [0:26] 观众被问及他们在场景中感知到了哪些变化。
- [1:39] 原始图像不同：建筑不是粉色的，顶部有窗户而不是门，栏杆是完整的，还有人和请愿标语。
- [1:56] 大多数观众完全没有注意到这些变化中的大部分。
- [2:03] 重放视频显示，即使知道变化正在发生，你也会看到一些但不是全部。
- [2:21] 眼睛并不直接看到世界，头部太小无法容纳它；眼睛每秒移动和注视约三次，并合成世界大致样子的图像。
- [2:42] 你看到的大部分实际上是想象出来的，想象力在本课程中扮演重要角色。
- [3:07] 编程一台能看的机器是一个严重问题，因为我们以为自己知道如何看，但演示表明这种观念是错误的，我们并不知道自己如何看。
- [3:27] 如果我们不知道自己如何看，那么在 Python 或 MATLAB 中编码并告诉机器去看是极其困难的，因为我们无法内省。
机器学习入门讲座
机器学习与手工编码视觉
- [4:55] 一个核心问题是世界很难向机器描述，例如定义长颈鹿很困难，因为长颈鹿在形状、背景和颜色上各不相同。
- [5:46] 即使编码像人脸检测这样简单的东西，手工完成也极其困难。
- [6:12] 观众中只有三四个学生知道如何构建人脸检测器。
- [6:33] 使用机器学习技术构建人脸检测器非常容易，但如果你尝试编写代码，则极其困难。
- [6:39] 日本的一家汽车制造商有一个手工编码的人脸检测器，它未能找到一位高个子金发朋友Jesse的脸，因为它是针对深色眉毛和眼睛编码的。
- [7:25] 手工编码的系统极其脆弱，仅在最佳条件下才能工作。
- [7:44] 机器学习解决方案是收集来自各种人群和群体的大量人脸图像，然后使用简单的算法来检测人脸。
- [8:03] 与其编写一个程序告诉机器什么是人脸，不如向机器展示哪些是人脸，哪些不是，并让它学习区分它们的特征。
- [8:26] 机器学习方法基于示例而非规则，让机器找出特征或模式。
训练机器检测人脸和行人
人类学习与机器学习
- [11:51] 人类和所有动物都是非常优秀的学习者，而人类尤其出色。
- [12:06] 一个观众练习测试人类从几个突出示例中学习“tufa”概念的能力。
- [12:25] 观众成员投票选出哪些指出的物体是tufa，有些获得多数同意，有些则很少。
- [13:21] 一位参与者通过茎、根和头来解释tufa。
- [13:59] 人们对判断tufa很有信心，尽管他们发现很难描述标准。
- [14:26] 人们从极少的例子中很快学会什么是tufa。
- [14:39] 关于什么是tufa的观点有些主观，但仍然存在强烈的一致性。
- [14:51] 像长茎、腿或头这样的特征定义了tufa，但特征提取主要是在潜意识中完成的。
- [15:19] 目标是设计能够从少量示例中快速学习新概念的机器。
- [15:32] 机器无法做出这些归纳跳跃，正如Josh和Yvonne所称。
- [15:42] 机器学习提取特征、模式或统计数据，使我们能够识别tufa、人脸或行人等事物，然后做出预测。
预测任务和机器学习应用
- [16:17] 预测可以像决策一样运作，例如决定某物是人脸还是不是人脸。
- [16:26] 预测也可以具有不同的性质，例如判断某物是否危险以及是否要逃跑。
- [16:46] 预测是一项流行的预测任务，例如预测建筑物的能源需求以调整你购买多少能源。
- [17:11] 销售预测是另一个例子，例如预测像粉色Lululemon产品这样的产品的需求。
- [17:32] 机器学习可以自动为你未评分的电影评分，以帮助构建更好的推荐系统。
- [17:50] 机器学习可以学习什么是HIV病毒，然后检测突变，这很重要，因为检测突变可以让你选择不同的治疗方法。
- [18:22] 信用评分系统快速分类你是否信用良好，这就是为什么如今使用信用卡如此容易。
- [18:49] 机器最终可能在乳房X光检查等医学诊断方面比人类做得更好。
- [19:01] 排序是机器学习非常有用的另一项任务。
图像搜索、摘要工具和作为AI的机器学习
- [19:07] 在谷歌图片搜索中点击图片会告诉谷歌该图片与搜索词匹配。
- [19:21] 谷歌图片搜索随着时间的推移而改进，因为人类通过点击来标记图像。
- [19:30] 搜索机器学习术语时，前50个结果中经常出现热门模特的图片，因为那是人们点击的地方。
- [19:40] 这些算法失败让演讲者意识到发生了什么。
- [20:00] 在当今巨大的信息流中，能够总结信息也很重要，许多工具已经尝试浓缩新闻或其他信息。
- [20:18] Zite由多年前参加这门课程的两名学生与BC省的一些风险投资一起构建，最终被CNN收购。
- [20:35] 为CNN提供支持的算法就是在这里开发的。
- [20:48] 机器学习是真正的AI，或者肯定是构建智能机器所需的AI的重要组成部分。
何时需要机器学习
- [21:10] 当我们缺乏专业知识时，就需要机器学习。
- [21:17] 火星机器人无法远程控制，因为信号传输时间太长，因此它必须了解其环境并自我修复。
- [21:29] 我们无法理解我们如何看到、识别事物，或从声音中看到概念，这种缺乏内省能力促使我们使用学习。
- [21:53] 学习对于不断变化的问题很有用，比如网页推荐，因为人们的兴趣会随时间改变。
- [22:10] 依赖策展人需要很大的勇气，所以自动系统对于随着兴趣变化持续提供内容很重要。
- [22:33] 因为每个人都不一样，一百万客户需要一百万个自动推荐器，而不是一百万个策展人。
- [23:04] 学习也适用于人类无法推理的问题，在搜索引擎或网络分析中你不能依赖规则。
数据增长与对机器学习的依赖
语言技术成功案例
语音识别与众包数据
数据作为关键商品
引文匹配与流行度搜索
信息抽取与网络真相的局限
- [33:06] 信息抽取被提出作为课程项目中值得追求的主题。
- [33:19] 展示的信息抽取引擎由西雅图华盛顿大学的教授Oren Ezioni和Dan Weld开发。
- [33:53] 目标是语义答案，例如谁与奥巴马结婚，而不是仅仅包含查询词的页面。
- [33:59] 该引擎通过从网络中提取名词-动词-名词三元组来工作。
- [34:19] 网络包含真相但也充满垃圾，仅使用计数或流行度来可靠地提取信息是极其困难的。
- [34:38] 当前的方法使用一堆规则，这被描述为当今可用的最佳方法，也是谷歌和许多公司使用的那种。
- [35:07] 超越当前规则有两条途径，一是参数化语言学家目前创建的规则。
- [35:34] 学习机器可以估计最佳预测标签，人们通过应用程序对标签投赞成/反对票。
- [35:43] 揭示真相最终需要应对因果关系并进行实验或行动。
- [36:14] 行动被认为是获得真理所必需的。
通过干预和因果关系进行学习
神经网络与深度学习
大脑、记忆与深度学习
- [37:59] 机器学习成为重大新闻，包括谷歌的神经网络从YouTube视频中学会了识别猫和面孔。
- [38:29] 过去三年深度学习取得了惊人进展，使我们更接近主动通用物体识别。
- [38:44] 微软和谷歌的语音识别因深度学习而改进，这个受生物学启发的主题将占课程约三分之一。
- [39:20] 进入眼睛的视觉信息微小且动态，高分辨率部分大约拇指大小，低分辨率部分大约手掌大小。
- [39:54] 视觉信号通过视神经传到大脑后部，那里的LGN大约有一百万条纤维，提供大约一千乘一千的分辨率，而V1区非常有趣。
- [40:30] 大脑是一个良好的记忆和预测装置，以联想方式存储事物，看到有云的飞机时能检索出飞机。
- [41:29] 一名学生说他的电话号码倒着背明显更慢，因为人们存储模式和序列，而不是像计算机RAM那样存储数字。
- [42:14] 一位非英语母语者说他用中文说电话号码很快，但用英语较慢，表明他在英语中读取数字的图像。
- [42:53] 联想可以是空间的也可以是时间的；擅长记忆的人会为物体构成图像并背诵该图像，创建模式和上下文。
蜘蛛认知与位置细胞
- [43:37] 关键点是记忆是联想的。
- [43:45] 观众被要求选出哪张图片是跳蛛，并按图片统计票数。
- [44:45] 这些图片被展示给一只雄性跳蛛，看它是否会将其中的一张视为想要进行蜘蛛性行为的雌性。
- [45:07] 即使是小动物，在识别我们的能力上与我们相差不远。
- [45:15] 蜘蛛能识别简单的漫画和图画，而不仅仅是真实图像。
- [45:28] 跳蛛甚至能从漫画中识别交配行为。
- [45:35] 跳蛛能识别3D深度，这对于在世界上移动至关重要。
- [45:51] 蜘蛛非常低级的大脑是一个极其精密的机械装置，我们仍然难以用电子设备复制。
- [46:16] 一些生物，比如这只蛾，进化出看起来像跳蛛的样子，以免成为它们的猎物。
- [46:27] 我们看到的大部分内容都被编码在大脑中，尤其是事物在世界中的位置。
- [46:41] 监测神经元是否放电显示其在盒子中放电的特定周期间隔，这就是你如何知道自己所在位置的方式。
V1神经元与分层特征学习
- [47:17] Hubel和Wiesel用垂直和水平条探测猫大脑后部的V1细胞。
- [47:56] 神经元不会对水平条放电，但只有当条处于特定角度时才会放电。
- [48:10] 每个V1神经元对特定的线条方向敏感，因此不同的神经元对不同的边缘放电。
- [48:35] V1中大约有20百万个神经元，每个监测一个斑块、坐标或边缘。
- [48:50] 这些基本特征的组合允许识别某物是否是脸。
- [49:27] 基本思想是每个神经元对一种特定刺激敏感，例如图像某部分的边缘。
- [49:36] 当神经元看到明暗对比时，它会放电并产生尖峰，否则不会放电。
- [50:43] 深度学习将这些神经元的输出馈送到第二个网络并学习其参数，产生诸如主体部分和整个面部的东西。
- [51:22] 谷歌在YouTube图像上无监督地训练了一个大型网络，能够识别面部的神经元出现了。
- [52:28] 这在生物学上也是合理的，呼应了神经科学中著名的哈莉·贝瑞神经元故事。
神经元激活与输入重建
语音识别流程与下一节课
车站容量与结束语
[0:00] 我要播放这个，然后按照顶部这里的说明操作。谢谢。谢谢。谢谢。谢谢。谢谢。谢谢。谢谢。谢谢。好，那么你们感知到了哪些变化——门的颜色更深了。
[0:36] 以这个速度。什么？我不会大喊大叫的。哦，对，那个标牌，参与者的标牌变了。谢谢。还变了什么？这是给所有人的，340。谢谢。什么？栏杆消失了。谢谢。所以特别是——原来有——这条红色链接消失了。那个标牌，是那个吗？谢谢。
[1:09] 右边的那栋建筑。有多少人认为右边这栋建筑与最初的那栋完全不同？请说。有340个人。好。有多少人认为这里前面有一只狗？在这扇门上方。有三四个人。其实没有。我只是在逗你们。那里有一个人。这是原始图像。所以如你们所见，那栋建筑并不是粉色的。它是另一栋建筑。
[1:43] 顶部是窗户而不是门。栏杆是完整的。那里有人，还有一份请愿书签名，而——这一切都在你们眼前发生了变化。你们中的大多数人对其中大部分变化完全没有察觉。我要重新开始，再播放一遍。只是为了说明这一点：即使你知道这些变化正在发生，你仍然无法把它们全部看到。
[2:18] 你会看到一些，但不会全部看到。这是因为你的眼睛并不直接看见世界。你的头脑里并没有装着整个世界。你的头脑太小了，世界无法把它填满。所以你实际做的是：你的眼睛在移动，每秒大约三次注视世界，然后用这些注视拼凑出一幅关于世界是什么样子的图像。而你所看到的大部分内容实际上是想像出来的。想像力在本课程中扮演着重要角色。本课程在很大程度上讲的就是想像力。
[2:50] 如果你能想像这个世界，那就意味着你已经学会了这个世界是什么样子。这些视频还说明了一件重要的事，那就是——如果你想给一台机器编程让它能看见，我们在这里面临一个严重的问题。因为我们以为自己知道我们如何看见、看见了什么。然而我刚才已经向你们证明，这种看法实际上是错误的。
[3:23] 我们实际上并不知道自己是如何看见的。既然我们不知道自己是如何看见的，我们又怎么能用代码、用Python或MATLAB来编程，告诉一台机器去看呢？这极其困难，因为我们无法内省。而如果我们无法内省，复制一个大脑就非常困难。谢谢。好。这就是——他对今天主题的简短介绍。
[3:57] 这就是你们的——第一讲。在今天的课上，我会——简要介绍什么是机器学习，以及可以在哪些地方使用机器学习。我会谈谈所谓的大数据革命，为什么机器学习突然变得……这与我们现在能够获取海量数据、拥有处理这些数据的计算能力有关。谢谢。嗯，我会谈谈课程中的一、一部分——会成为这门课重要部分的内容：神经系统，以及我们如何从生物系统中获得灵感，构建能够学习的机器。
[4:40] 最后，我还会——简单谈几个应用。希望这已经能给你们一些关于项目该做什么的动力。好。那么我们如何告诉机器如何看见东西呢？所以一个大问题是，世界并不容易被描述。如果你想告诉一台机器如何识别长颈鹿，那会非常困难，因为长颈鹿有各种各样的形状。可能还不至于像这里这些被夸张了的长颈鹿那么糟。
[5:17] 但很难告诉计算机长颈鹿是什么。因为它们的形状五花八门，或者出现在不同的背景中。你可能会在非洲看到长颈鹿，但突然间，你又会看到雪地里的长颈鹿。很快在卡尔加里。而且它可能是白化长颈鹿，那样的话，颜色就能帮上忙。嗯。那么我们如何告诉——我们如何编码——什么是一头长颈鹿？事实证明这极其困难。或者一个更简单的问题，我们如何编码——什么是一张脸？我们如何告诉一台机器那张图像里有一张脸。
[5:54] 而这算是——你习以为常的东西，因为——你所有的iPhone或手机都有面部检测器。如今的相机都带有FaceTime检测器。你用Facebook，它有面部检测器。它知道图像中脸在哪里。那么。你们中有多少人知道如何构建一个面部检测器？只有三四个学生举手。你们见过随机森林。别泄气。也许我们会把这作为本课程的一次作业练习。
[6:28] 构建一个Facebook。用机器学习技术非常容易。但如果你试图通过编写代码来实现，那就极其困难。所以我总是告诉人们的一个轶事是：有一次我在为日本的一家汽车制造商做咨询，当我到那里的工厂参观时，我坐在车里。它立刻就知道我的脸在哪里。所以有一个摄像头，还有意图，这样他们就知道我在做什么、在看什么，以及我是否在打瞌睡。
[6:59] 作战。茶货上。于是，但是后来，我的一个朋友，Jesse，他现在是滑铁卢的教授，Jesse又高又是金发。而这个东西找不到他的脸。原因是因为这里面只有日本工程师。他们编码的脸基本上就是深色眉毛、眼睛等等。所以这个系统，这个手持摄像系统极其脆弱。所以这就像一个被编码来识别你的操作的系统。
[7:32] 以及最优条件。但当你逃离捕食者奔跑时，却不可能识别出你出局了。再见。机器学习解决方案是——收集大量的人脸图像。这些将来自各种不同类型的人、人群等等。然后想出一个非常简单的算法，让你能够检测人脸。所以你不能告诉机器——怎么做？
[8:06] 你无法编写一个程序来确切地告诉机器什么是脸，但你可以编写一个程序，向机器解释这些是脸。这些不是脸。然后让机器自动学习是什么区分了脸和非脸。这就是机器学习方法。它基于示例，而不是我们告诉规则。我们让机器弄清楚哪些是东西，哪些是特征，我们称之为特征或模式，或统计数据，使脸不同于任何其他图像。
[8:47] 所以第一部分是收集那个巨大的人脸图像集合。然后你收集一个已知人脸的大数据集。而——结果就是这个。这就是你的人脸检测器是如何构建的。嗯。我将在课程中向你展示如何用一种叫做随机森林的技术来做到这一点。这样你就能拥有漂亮的消费电子产品等等。嗯。当然，这不仅仅是用于人脸。你可以对任何东西、任何物体做这个。所以你可能会拍摄大量行人的图像，并训练机器识别行人。
[9:24] 现在，我为什么想要机器识别行人？谢谢。你是什么意思？自动驾驶汽车。自动驾驶汽车不远了。它们已经在加利福尼亚获批。嗯，而且，嗯……很快。呃……希望街上会有自动驾驶汽车。嗯……我试图回忆一个数字，二十或四十，要么是20要么是40。
[9:56] 我把我的力量设为困惑。每年在美国有20,000或40,000人死于车祸。大多是行人。为了把这个项目的重要性放在背景中，想象一下UBC的人口每年消失。这就是不正确地驾驶并依赖人类的代价。试试看。在印度，几年前我访问时查过，是113,000人。
[10:31] 那个职业，这可能是好事也可能是坏事，取决于你怎么算。现在，嗯……实际上，无论你怎么算，它都是坏事。113,000上升。巨大数量的人。谢谢。所以——同样的想法。*咳嗽*你教机器，这里有一堆图像，在这种情况下是汽车的图像，然后再次，如今的商业系统，比如Mobile已经能够检测火焰、检测汽车等等。
[11:06] 所以——但所有这些巨大创新的关键，能够拥有一个自动的——因为如果没有这些创新，根本不可能拥有自动驾驶汽车。因为你怎么能部署一辆看不见并理解前方行人是什么的汽车呢。那就像是谋杀。嗯……关键是要意识到我们可以教机器它们所接受的东西。所以这门课确实将关于如何编写程序，让机器抓取数据并。
[11:42] 决定什么是脸，和/或什么是行人，等等。嗯。现在，外面还有其他学习机器，而那些学习机器就是人类。人类非常擅长。事实上，不仅是人类，所有动物都是非常好的学习者。嗯。人类尤其擅长。嗯。我现在要用这个例子测试人类学习得有多好。所以，我在这里高亮显示了，我有一堆图像。
[12:15] 物体的。其中三个我称之为tufas。好的，你们没有通过给它们加红框来标记它们。所以现在你们的问题是，你们中有多少人认为我箭头指的这个是tufa？好的。大多数。你们中有多少人认为这是two-fly？这个呢？那是怀疑者。
[12:48] 这个呢？那是tufa吗？只有一个？所以来看一个有趣的样本，这个。谢谢。好的，不，没有那么多人。嗯……这个家伙。——这个家伙？亲爱的各位，是的。好的，让我们选一个喜欢的人，让我们从那个开始。
[13:21] 为什么那是tufa？上帝保佑你。茎，而且它有点根。还有好的，茎、根和一个头。这就是使它成为tufa的原因。呃……还有谁想评论为什么他们说某物是twofer？或者不是。——看着狗，有一个长皱眉。有一个松的。
[13:52] 只是有点像红牛。谢谢。来吧。很难描述，但不知怎么你就是知道。因为你——非常自信地举手或把手放下。即使你无法描述它，说那是tufa或不是tufa。看到了吗？你知道什么是two-finger吗？谢谢。大多数人。还有其他人吗？
[14:23] 谢谢。谢谢。所以……这说明了几个事情。一是你学习什么是two-fay有多快。你从未见过two-foot的人，也许在你的数据领域里，但你很快就学会了什么是2P。谢谢。从很少的例子中。嗯……并非所有人对什么是truffe都有相同的看法，这就是为什么我问了几个问题。所以它有些主观。但尽管如此，还是有强烈的一致性。嗯……然后……有一些特征使twofer成为twofer，无论是长茎还是腿头。嗯……
[14:59] *咳嗽* 但总的来说，当我们有意识地提取这些特征时，我们甚至不会想到它们。这是潜意识在起作用。很多特征提取和对世界的解释都是通过潜意识完成的。潜意识非常强大。也许比意识更强大。嗯。我们希望设计出能够做到这一点的机器。它们能看到几个数据示例，并且仅凭几个示例，就能非常快速地学习新概念。
[15:32] 机器无法做出这些归纳。飞跃。正如乔什和伊冯娜所称的那样。所以机器学习是关于能够提取特征、模式或统计数据。让我们判断什么是双折的东西，一旦你有了这些特征，你可能不会自觉地这样做。我们人类不会自觉地这样做，这就是为什么很难。试图手工编写代码，而我们必须依赖机器。
[16:07] 是提取构成双折的模式。构成一张脸或一个行人的模式，然后一旦你有了这些，就做出预测。预测可以像决策一样。决定某物是不是一张脸。某物是不是一张脸。或者不同性质的预测，比如，这东西会吃我吗？我应该逃跑吗？这太危险了吗？谢谢。
[16:39] 还有其他类型的预测任务现在非常流行，所以我快速介绍其中一些。嗯。预测。嗯……这个我会在下一节课作为例子来做。我们经常用它来——预测你的建筑需要多少能源需求。因为基于此你可以调整嗯……你知道，你购买多少能源。而且——你知道，可能节省。嗯……为经济节省开支。
[17:10] 谢谢。你可能想预测销售额。如果你在Lululemon工作，你会预测有多少人会去买粉色的那个——在Taskware。机器学习实际上非常有用。当然，GAAP等北美各地的商店非常——一系列其他机器学习。谢谢。你可能想——你可能想——你可能想你可能已经给几部电影评了分，所以你会希望机器自动给其余的电影评分，以保护你会给任何其他电影分配多少颗星，这样他们就能构建更好的推荐系统。
[17:48] 嗯……你可能想使用机器学习，一旦你教会机器学习什么是HIV病毒，你希望机器能够检测到它所监测的东西中任何偏离病毒的情况。换句话说，你希望机器检测突变。这些非常重要，因为一旦你检测到突变，你就可以选择不同的治疗方法，呃……治疗病毒。
[18:21] 嗯。如今使用信用卡如此简单的原因是因为有非常好的信用评分系统。所以机器学习会快速分类。将决定你是否——呃……你是否通过测试，你是否信用良好。所以我们经济的很大一部分已经——依赖于使用非常大规模经典的现金。以及医疗你知道，诊断。
[18:51] 比如乳房X光片等，机器最终会比人类做得更好。排名，另一个机器学习非常有用的任务。所以如果你使用谷歌图片搜索，例如，你经常做一个搜索。然后你点击一张图片。当你这样做时，你基本上是在告诉谷歌，嘿谷歌，对于这个词，这是与之匹配的图片。
[19:21] 这就是谷歌图片搜索如何随着时间的推移变得越来越好的原因，因为我们——我们所有人类都在通过点击来标记图片。通常你搜索一个机器学习术语。在顶部，显示的前50张图片，有这些性感模特的照片。那是因为当人们搜索支持向量机或他们想要的任何技术时，他们点击的就是那里。所以算法有一些失败，这让我意识到他们在做什么。
[19:54] 谢谢。呃……当然——嗯……在我们今天生活的这个巨大的信息流中，能够总结信息也很重要。已经有很多工具会尝试为你浓缩信息，无论是新闻还是任何其他类型的信息，并使其可用。其中一个工具是——嗯，是Zite，它是由两个多年前上过这门课的学生，我的课和Eric Berchu，构建的。
[20:26] 他们，连同BC省的一些风险投资，最终建立了这家公司，后来卖给了CNN。如果你去CNN网站，驱动CNN的算法就是在这里开发的。好的，所以……最后嗯。经常谈论AI时，我认为机器学习才是真正的AI。当然是AI的一个巨大组成部分。
[20:58] 或者构建智能机器所需的东西，但它当然是一个巨大的组成部分，而且在过去十年中取得了巨大进展。非常感谢，Imran。谢谢。现在，我们什么时候需要机器学习？当我们的专业知识缺失时。如果你想送一个机器人去火星，嗯……远程控制行不通，因为传输信号需要太长时间。所以你需要一个能理解环境、能自我修复等的机器人。谢谢。嗯……从视频中可以清楚地看出，我们无法理解我们如何看到或识别。
[21:34] 事物，甚至我们甚至不理解我们如何——从声音中看到概念。因此——缺乏内省能力促使了学习的运用。嗯……当你处理不断变化的问题时，这也很有用。比如，如果你试图给某人提供好的推荐，我不知道，猫咪或狗狗，或者现在人们在网上推荐的那些东西。
[22:06] 嗯。人们的兴趣随时间变化。所以如果你依赖策展人，嗯……那将——那需要很大的勇气。所以重要的是，嗯。要自动化，拥有自动系统，这样随着人们的兴趣随时间变化，你仍然能向他们提供他们需要的内容。嗯。与此相关的是这个问题：通常每个人都是不同的。如果我想推荐新闻，我不必为每个不同的人构建一个新闻推荐系统。
[22:41] 因为每个人……这意味着，如果我有一百万个客户，我将需要，比如说——与其拥有一百万个策展人，我想要的是一百万个自动推荐内容。嗯……当然——嗯。对于人类无法推理的问题。当你在搜索引擎公司、在做网络分析的公司工作时，很快就会明白，你不能依赖规则——你不能依赖人去做事。问题的规模如此之大，你确实必须依赖机器学习。
[23:31] 现在说数据。过去几年里，数据大幅增加。嗯，正如——Lon、Halaby、Peter Norbeck和Fernando Pereira，Google的研究人员，Google语言领域的几位顶尖研究者。嗯。在论文中指出的那样。在2009年。在1960年，嗯。一百万词曾被视为一个大型数据集。在2006年，我甚至不知道那个数字是多少。
[24:03] 嗯——万亿。就是这样。谢谢。呃……谢谢。词数据集在学术界曾被视为大型数据集。所以已经出现了巨大的数量级。在你们能够在个人对话中实际管理的数据集上。然后如果你去Amazon EC2等等，就变得更有趣了。语言的成功故事是什么。
[24:37] 我出局了。在过去几年里。技术上的成功。Watson。什么？那是什么？Watson？Jeffrey Lee Chan。可爱。成功故事。成功故事是——而且它们实际上为很多人创造了收入，解决了人们的问题。Google Translate。ERIC SCHMIDT：但Watson是个好例子。Google Translate。那是个绝妙的例子。因为……嗯。
[25:07] 我妻子可以在Aritzia做她的工作，她需要来自印度的纺织品，她需要有人在中国编织这些纺织品，她给印度、中国的所有这些人发邮件。他们可能说印地语，可能说普通话，而他们都能沟通，因为他们都在使用Google Translate。所以它彻底改变了电子商务。因为现在世界变得更小了。你甚至不需要都说同一种语言就能沟通。
[25:40] 偶尔，这会有一些问题。这基本上就是人们做事的方式。即使在温哥华的小公司里，这就是——这就是——这就是我们公司能够用他们的产品参与全球业务的方式。另一个例子是什么？Apple大力营销了它。语音识别。所以，是图像识别理论和——所以——所以。许多粒子。
[26:12] 我们在加拿大。但是——Google的。语音识别效果非常好。我不知道你最近有没有试过。如果手机上的语音识别。美妙之处还在于，每次你弄对或弄错，你都在为Google提供反馈。变得更好。所以我假设语音识别是那些已解决的问题之一。嗯。你如何——谢谢。
[26:45] 年度教师联合。欢呼。这很奇怪。按我的书来说。什么？类似实时翻译。成另一种语言。而且……哦，是的。你甚至可以两者都做。所以最终我们将能够说话，如果你愿意的话。是的，就像科幻小说。在那里我们都能……也许我会轻触说话。父亲。我不会感到惊讶。
[27:17] 就使用深度网络的语音识别而言，嗯。这是我稍后课程中要讲的内容。有两个团队在做这件事。第一个开始做的团队实际上在soft。然后Google也开始做了。所以他们使用的是稍微相同的——那种基础设施。关于那个。所有这些的共同点？是有大量数据。有两种语言的大量数据。而且有——大量——换句话说，匹配数据。所以这是法语文本，这是翻译。
[27:55] 嗯。语音识别也是如此。人们创建了这些巨大的数据集。此外，数据集每时每刻都在变得更大，因为像Google和Microsoft等公司一直很出色的地方在于，他们给你一个服务，如果你不喜欢翻译，他们允许你纠正它。而现在正因为如此，你一直在为这些公司提供数据，让他们在提供产品方面变得更好。所以他们在利用群体智慧。
[28:26] 谢谢。呃……现在，对于很多这些应用，你确实需要大量数据。大量数据是必不可少的。如果你要做语音识别，真正的问题甚至不是知道如何构建技术。嗯……而是实际上拥有数据。来训练技术。数据是商品。是的。这里有一个例子来说明这一点。由CMU的Alyosha Efers。所以他最初有一张他不太喜欢的图像，因为里面有一栋房子。所以他想要的是——我能移除这栋房子吗？
[29:05] 所以你显然可以把它移除。然后问题是，如果你有一组来自网络的其他图像，你能用这另一组图像来填补这里吗？你知道的。以某种看起来比这个更好的方式。如果——他发现如果他只用几张图像，效果不好。但后来当他增加这个数量时，她能够让机器自动合成新的图像，那——看起来仍然非常可信。要实现这一点，你确实需要大量数据。
[29:42] 现在，回到谷歌那些人，他们在论文中说——他们已经解决了如何让人们贡献数据的问题。而且——当然，Facebook和谷歌在这方面做得非常出色。嗯……Dave？他们声称已经解决了聚合和索引数据的技术问题。他们在某种程度上做到了这一点，谷歌当然做得很好。还有很多工作要做。
[30:14] 但最重要的是，我们仍然没有解释内容。所以这里有一个例子，引文匹配。现在，当你在谷歌中进行搜索时，你本质上使用倒排索引，你在搜索中查找那个词，检查那个确切的词是否出现在网页中，然后返回那些网页。然后还有一些使用算法完成的排名——如果你做Python网站上的Python练习，那将是学习如何编写谷歌搜索引擎的一种方式。
[31:04] 现在……然而，以这种方式进行搜索的问题在于，它非常基于流行度。谷歌是一台流行度机器。现在，一个反例是引文匹配。所以引文匹配是试图判断两条引文是否相同的问题。所以在科学论文中，你总是在最后引用别人的工作。所以这是你在为项目写报告时会做的事情。
[31:38] 判断两条引文是否相同实际上被证明是极其困难的。这是因为人们经常拼错名字。人们经常写卷号或不写，包括页码，或者页码错误。而学术界经常发生的情况是，如果我引用了一部作品而我没有把页码写对，然后另一个人想引用我的作品，但我的作品实际上追溯到另一篇论文。他们并没有读过那篇论文，但他们仍然引用了它。而他们所做的是复制引文。
[32:14] 而且你也只是在截止日期前几个小时，你必须提交一篇只是剪切粘贴了一堆引文的论文。当最后交项目时，你会明白我的意思。所以这样做的问题在于，如果某处有错误，它很容易传播。现在你有898条页码错误的引文和22条页码正确的引文。如果你只按流行度来，你会选择那些。所以你必须做别的事情。你实际上必须进行干预才能解决那个问题。
[32:54] 让我给你看另一个例子。嗯，这是信息抽取的问题。我很想在本课程中看到几个关于这个主题的项目。所以这个信息抽取引擎是由Oren Ezioni和Dan Weld开发的，他们是西雅图华盛顿大学的教授。
[33:26] 所以这让你可以输入像奥巴马。结婚。这样的东西。现在我们开始做一些更有趣的事情，一些具有语义价值的事情，而不仅仅是一个查询，比如，你知道的，不仅仅是——输入像奥巴马结婚并检索一堆说奥巴马结婚的页面，而是你实际上想知道谁和奥巴马结婚了，或者他和谁结婚了，如果你进行搜索，他们通过从网络中抽取名词、动词、名词的三元组来做到这一点。
[34:08] 所以这就是你得到的结果，我想象的数字是频率。那就是网络。网络有真相，但网络也充满了垃圾。仅仅通过使用计数和流行度来可靠地从网络中抽取信息是极其困难的。而你能解决这个问题的唯一方法就是进行干预。嗯，有两种方法可以做到这一点。目前，他们使用一堆规则。顺便说一下，这是最好的。
[34:44] 这就是精华，你在这里看到的是我们今天能做的事情。这是谷歌会使用的那类东西。等等。很可能，很多公司都在使用它。Orin Asuncion是一位伟大的企业家，他使用这类技术卖掉了许多公司。如果我们要超越这一点，有两条路线。一是开始将这些规则参数化。因为现在，语言学家倾向于创建所有这些句法规则等等。
[35:19] 它必须是一个名词短语后跟一个动词短语后跟一个名词短语，等等。或者然后他们开始添加一堆例外。如果你能把所有这些都参数化，然后让一个学习机器估计什么是最好的、预测标签，然后让人们投票赞成/反对标签。如果你能创建一个很酷的应用让人们这样做，我们就能学会做信息抽取。好得多。但如果你想揭示真相，最终我们需要应对因果关系的概念。我们需要真正知道什么是……我们需要进行实验。
[35:56] 或者我们需要真正去探查，我们需要采取行动才能实际弄清楚。我们必须在现在，在这个点之后做些什么。我们将不得不点击，我们将不得不做一些研究，我们将不得不决定我们将如何进行那项研究，以便能够发现真相。所以行动对于真相是必要的。这把我带到下一部分。
[36:38]也就是——学习并不是仅仅在有大量数据的情况下发生，然后你从中学习。我在340里就是这么做的，为了那里的那些人。但在540中，我们还会考察我们何时真正主动去寻求数据，何时进行干预。因为大多数系统并不仅仅是……被数据淹没并据此做出反应，而是你实际上要在世界上进行干预。而且你知道，大多数数据，为了获取数据我们要做实验。我的意思是，这基本上就是科学方法的核心。
[37:20]如果我们想理解某件事，就需要采取行动。所以，今年这门课我会讲一点因果性，因为我认为这很重要，最终是为了获得真正的理解。好。这门课还会涉及大量神经网络。最近，在一个叫做深度学习的领域里出现了很多很酷的东西。你们中有多少人看到了《纽约时报》上关于谷歌Frontline深度学习网络的报道？
[37:59]一，看来机器学习在新闻里很火。火到有两个人、三个人、四个人、五个人看到了它。它登上了《纽约时报》的头版。所以谷歌声称他们构建了这个神经网络，它能查看YouTube视频，仅仅通过这样做就能学会什么是猫、什么是人脸等等。所以这门课我们会仔细讲解那个实验。
[38:29]深度学习的最高水平。但不用说，过去三年里这个领域取得了一些真正惊人的进展，使我们更接近于实现主动的通用物体识别。语音识别在过去——在微软和谷歌——变得好多了，它变好的原因就是这种深度学习。所以这门课有相当一部分内容，而且它都非常受生物学启发。所以我们会在课程中涵盖这些。
[39:03]这几乎会占课程的三分之一。我不会深入讲神经科学，但我不会停留在屏幕上这个小漫画上，也就是这里有一个大脑。如果你在看视觉信息，会有一个视觉信号进入你的眼睛。再说一次，那个视觉信号非常小，而且是动态的，它大约有你——高分辨率的部分大约有你拇指那么大。
[39:37]——在这个距离上。然后较低分辨率的部分大约有你手掌那么大。但实际上你看到的并不多。这其实大部分是想象出来的。然后有——一切都经过——这条视神经传到大脑后部，那是视觉的第一部分，试试这个，试着敲一下你的后脑勺。眼前会变黑。这是有原因的。你的视觉就在那里。
[40:10]这边有LGNAs，大约有一百万根纤维。所以这开始让人大致了解那幅图像的分辨率，一千乘一千。然后在后部那里，有一个叫做V1的区域，非常有趣。现在我得对此说几句。大脑是一个很好的记忆装置，也是一个预测装置。我们在大脑中存储事物的方式是联想式的。
[40:45]我的意思是，比如说，像那样，那里有——如果你想检索某样东西，当你提出一个新的查询时，比如在这个例子中，一架有云的飞机，你就检索出一架飞机。所以它理解那里有一架飞机。嗯，谁不介意把手机借出来？你能大声把你的电话号码正着说一遍吗？604-442-7185。
[41:19]你能倒着说一遍吗？我听到的是280。——好，所以他倒着说的时候明显慢得多。原因是你并不像计算机那样处理数字。你并不像，你知道的，用随机存取存储器之类的东西来存储数字。你学到的是你重复的这个模式，也就是7782391190。所以很容易，因为就像这样，你存储了那个模式，也就是你的电话号码。你并没有存储那些数字。
[41:55]存储的是一个模式、那个序列，字母表也是如此。很多人正着说字母表毫无问题，但人们倒着说字母表就很吃力，因为他们学到并存储的是模式，好吧，序列等等，所以我有不同的体验，我不是英语母语者，但当我用中文告诉别人我的电话号码时，相当快。所以当我用英语说的时候，需要一些时间。
[42:33]所以我的感觉是，当我用英语说我的电话号码时，我……浮现出我电话号码的画面，然后读那个画面。我认为那可能就是时间花在哪里的部分原因。是的，确实如此。但你是在通过联想来做。所以联想，顺便说一下，这里我是用时间做联想，但它也可以是空间的。事实上，很多非常擅长记忆的人，是因为他们想到一幅图像。
[43:07]比如如果我给你一串100个物品，那么他们尝试做的，100其实是大材小用，10个物品。所以大多数人会觉得记住10个物品真的很难。但如果我，那些擅长的人，他们会用这10个物品组成一幅图像。然后他们只是背诵那幅图像。但他们创造了一个模式和上下文，这使它更容易。跟你说的有点类似。关键在于记忆是联想性的。
[43:41]一个游戏，今天玩的。如果你要从这些中选一个作为蜘蛛，你会选哪个？一只跳蛛。你们中有多少人见过跳蛛？它们到处都是。那么有多少人会投票认为第一个家伙是跳蛛？你只能选一个，所以这个有两票。这个。零。下一个。
[44:12]零。下一个。三票，四。下一个。四。下一个。一，三。这个家伙。一票，二，三。下一个家伙。好，那么这就像是，那30呢？好。现在……
[44:45] 如果我们比较这些数字，那么那些是展示给雄性跳蛛的图像，以观察这只雄性跳蛛是否会将其视为想要进行蜘蛛性行为的雌性。你绝对没想到我会给你看这个。所以——而我想在这里说明的一点是，即使是小动物，在识别我们的能力方面也与我们相差不远。
[45:15] 所以蜘蛛实际上能够识别简单的漫画、图画。所以不仅仅是真实的图像——那是一只跳蛛。而在这里你可以看到那些跳蛛能够识别交配行为等等，即使只是漫画等等。所以蜘蛛，这些跳蛛，它们能够识别3D深度等等，这对于在世界中移动至关重要。所以，我们不需要达到高层次的大脑。
[45:48] 我从高层次的大脑开始，它们展示的是人类大脑。但蜘蛛非常低层次的大脑是一台极其精密的机器，我们仍然在努力用我们自己的电子设备来复制它。而进化是美丽的，就像蜘蛛学会了所有这些惊人的——呃……心理特征、识别等等。有些生物已经进化出来，比如这只蛾，它们进化得看起来像跳蛛，以便它们不会成为它们的猎物。
[46:26] 嗯……我们在世界上看到的大部分东西都会被编码到我们的大脑中，但更重要的是，事物在世界中的位置。就像在这个例子中，我们用红色看到的东西。所以那是一个盒子。而那是正方形。还有一个筏子在盒子里移动。然后我们监测的是神经元是否放电。正如你所看到的，在盒子中有特定的部分、周期性的间隔，神经元会在那里放电。这就是你如何判断你在哪里的方式。闭上眼睛，就让那个神经元流动吧。
[46:58] 放电告诉我你在哪里。你在哪里被编码在你的大脑中。现在，回到视觉，回到这个，所有这些纤维进入这些第一层细胞。以及V1。人们做了同样的实验。在那里，探针。在猫身上。谢谢。在大脑后部。然后他们将这些条形如屏幕上所示进行短路。
[47:30] 竖条和横条。当你这样做时，嗯……这个实验是由Hubel和Wiesel做的。如果你在Google上搜索Hubel和Wiesel，你实际上可以在YouTube上看到整个视频。嗯——他们发现的是以下内容。如果——如果条形是垂直的——是水平的，抱歉——如第一幅图所示。如果条形是水平的——那么神经元不会放电。神经元只在条形处于特定角度时才放电。
[48:05] 角度就是神经元开始放电的地方。这意味着这些神经元中的每一个基本上都调谐于——垂直的、特定的线条。线条是不同的。方向。所以所有这些神经元基本上都有一个世界的年龄地图。因为它们放电——不同的神经元会为不同的边缘放电。现在那里有很多神经元。在V1中。这个拼贴中的每个小盒子都是那些神经元之一。而每一个都会——全部20百万左右。
[48:37] 将会监测检查这里是否有一块，那里是否有一个坐标。这里是否有一条边缘。而正是所有这些特征，不仅仅是眉毛。而是这20百万个特征，当你对它们进行组合时，使我们能够识别某物是一张脸还是不是一张脸。它们都非常基础。而这是——但通过拥有这些巨大的组合，我们能够进行各种识别。谢谢。
[49:08] 所以基本思想——我们将在课程中详细讲解这个，我只是……在这个阶段，我只是给出一个概述，希望能激励你继续下去。嗯。而在课程后面，我们将实际研究我们如何构建这些的细节。呃……但基本思想是你会有神经元，每个神经元都调谐于一个特定的刺激。就像在这个例子中，图像这个部分出现的一条边缘。谢谢。所以我们的本质基本上是明暗之间的对比。
[49:41] 而当那个神经元看到这个时，它就会放电。FIKE被产生。几个尖峰开始被产生。如果一个神经元没有看到这个，它就不会放电。而这是构建神经网络的基础。某种查看图像的东西，然后当你有2000个神经元连接到整个图像时，其中一些单元会放电，一些不会放电。这意味着我们正在将图像映射到一个代码。
[50:17] 大约2,000年。我们将在本课程中看到如何学习这些。如何学习，每个神经元的感受野是什么。每个神经元应该注意什么？什么年龄？所以我们将学习这个。对于这些神经元中的每一个，我们将从数据中自动学习它们。然后研究人员在深度学习中所做的是，他们将这些神经元的输出馈送到第二个网络。
[50:50] 在那里他们再次学习那个网络的参数。他们发现，当他们使用图像时，他们学到了诸如主体部分等等的东西。当他们只用，比如说，人脸图像来训练它时。然后在这里更往下，你实际上有完整的人脸。所以，所有这些如今都是自动完成的，而这本质上就是Google所做的——Google拿这些大型网络之一，能够直接训练它。而——而这里训练它的方式将是无监督的。
[51:27] 所以这会很有趣。我们不会说这些是人脸的图像，不是人脸的图像。我们将学习如何获取大量图像集合，并且仅通过使用一些概念——基本上就是你想象它应该看起来像什么、你看到的是什么。就是那个非常基本的想法。如果我有一个能够想象——嗯——它所看到的东西的学习系统，它就应该能够学习到有趣的概念。而这本质上就是 Google 所展示的。他们采用了一个系统，在其中你输入许多图像。说实话，这个想法此前已被学术界的许多团队所证明。
[52:06] 但 Google 只是把它提升到了下一个层次，非常大规模。于是他们——仅仅通过在 YouTube 上各种图像上进行训练，能够识别人脸的神经元就涌现了出来。所以在这组神经元中有一个特定的神经元。每当它看到这些面孔之一时，它就会激活。所以现在这也是生物学上合理的。有——论文。有一个非常著名的“哈莉·贝瑞神经元”的故事，神经科学实际上进行了一项研究，你知道，我用的是哪种动物？
[52:42] 从聊天中得到的答案。大概是猫。呃……人类是不允许参与这些实验的。嗯——那个会激活——有一个神经元，每当这只猫看到哈莉·贝瑞、看到哈莉·贝瑞的照片时就会激活。谢谢。谢谢。还有呃……谢谢。然后他们在实践中也可以处理。打开这个神经元，而且因为已经学习了感受野，你实际上可以学习，“什么样的输入会让这个神经元激活？”
[53:18] 而那就是登上《纽约时报》头版的那张图片。世纪。在课程的剩余部分，我们将看到如何计算这张图片。我还会看到，你可以重建输入这同一个想法，如何能被用来做这种——追踪，并能够补全这个所见。为了做出这个开放。保护。嗯……最后，作为结束，这是……如今你手机上能用的语音识别。它做第一层，就是从声音到倒谱系数。那算是非常简单的。
[53:59] ——二，二？标准。然后它使用深度学习，同样是神经网络，从那些基本上是声音特征的东西，到可以输入到一种叫做 HMM 的东西的特征，以便进行识别。所以贯穿整个课程，我们会去研究那些必要的元素。嗯——抱歉，发布。差不多就这些。所以下一节课将直接开始——开始——开始讲监督学习。
[54:30] 我们将从线性模型开始，因为它们恰好非常容易理解，我会让他拿着它。谢谢。再见。谢谢。谢谢。下次，我想要俄罗斯。
[55:10] 是的，是的。该站是一个大型多用途车站，容量约为 1,000 人。该站是一个大型多用途车站，容量约为 1,000 人。
[56:10] 谢谢。是的，我知道。
[57:10] 谢谢。*字幕由 Project readOn 提供* 谢谢。
[58:40] 我就坐在那里。我不认为我们处于联邦层面的状况。
[59:40] .
查看英文原始字幕
[0:00] I'm going to play this and then just follow the instructions here at the top. Thank you. Thank you. Thank you. Thank you. Thank you. Thank you. Thank you. Thank you. OK, so which changes did you-- perceived color of Taller of the doors.
[0:36] At this rate. Pardon? I'm not going to be screaming. Oh, yeah, the sign, the participates sign change. Thank you. What else changed? And this is for everyone, 340. Thank you. The what? The railing disappeared. Thank you. So in particular-- There was-- this red link has disappeared. the sign Is that friend? Thank you.
[1:09] The building on the right. How many people think that this building on the right is completely different than the initial Go ahead. There's 340 people. Okay. How many people think there was a dog here in front? Up this door. There's three, four people. There wasn't. I was just tricking you. There's a person there. This was the original image. So as you can see, the building wasn't pink. It was a different building.
[1:43] There were windows on top instead of doors. The rail was complete. there were people there was a petition signed And-- that all change in front of your eyes. and most of you were completely oblivious to most of these changes. I'm going to restart it and play it one more time. just to illustrate the point that even if you know that these changes are taking place. you will yet again not be able to see them all.
[2:18] You will see some, but you won't see them all. And it's because your eye does not see the world. You don't have the world in your head. Your head is too small. the world cannot fill it in. So what you do is your eye moves. and gazes at the world about three times second and then you compose with those cases an image of what the world is about. And most of what you see is actually imagined. And imagination plays a big part in this course. This course, to a large extent, is about imagination.
[2:50] If you can imagine the world. that means that you have learned. what the world is about. The videos also illustrate an important thing And it's that-- If you wanted to program a machine that sees We are confronting a serious problem here. Because we think we know how we see and what we see. And yet I've just proven to you that that notion is actually wrong.
[3:23] that we actually don't know how we see. see and now if we do not know how we see How can we code? and or Python or MATLAB. and tell a machine to see. comes extremely hard because we can't introspect And if we can't introspect, it's very hard to replicate a brain. Thank you. All right. That-- his minute introduction to the topic of today.
[3:57] which will be just your-- first lecture. And in today's lecture, I will-- tell you a bit about what is machine learning. And where you can use machine learning? I'll talk about the sort of the big data revolution, why machine learning has all of a sudden become and attested with the fact that we now have access to gazillions of data. computing power to process that data. Thank you. Um I'll talk about a a Part of the course-- something that's going to be a Big part of the course, neural systems, and how we draw inspiration from biological systems in order to build learning machines.
[4:40] And finally, I'll just-- touch on a few applications. which hopefully will start already giving you some motivation for what to do for your project. Okay. So how do we tell machines how to see? things So one of the big problems is that the world is not easy to describe. If you wanted to tell a machine how to recognize giraffe, that's going to be very hard because there's giraffe of all shapes. It might not be as bad as these giraffe here, which have been exaggerated.
[5:17] But it's very hard to tell a computer what a giraffe is. because they come in all sorts of shapes or in different backgrounds You could have a giraffe in Africa, but all of a sudden, you get a giraffe in the snow. Calgary soon. And it might be an albino giraffe, in which case, color would then help you. Um. So how do we tell--how do we code? what Azure Office is. It turns out to be extremely hard. or And even simpler problem, how do we code-- what a face How do we tell a machine There's a face in that image.
[5:54] And that's sort of-- something that you take for granted because-- all your iPhones or phones have face detector. Your cameras these days come with FaceTime. detector. You use Facebook, it has face detectors. It knows where the faces are. in the images. time. How many of you know how to build a face detector? There's just three, four students their hands up. You've seen random forests. Don't put fuel out. Maybe we'll do that in this course as a homework exercise.
[6:28] Build a Facebook. It's very easy with machine learning technology. But if you try to do it by coding it, it's extremely hard. So one anecdote that I always tell people is that, Once I was doing-- consulting for a car manufacturer in Japan and when I sat I was doing a visit there to the factory And I sat at the car. And it immediately knew where my face is. So there's a camera. and intent so that they know what I'm doing, what I'm looking. and whether I'm falling asleep.
[6:59] warfare. tea goods on. And so, but then, A friend of mine, Jesse, who's now a professor in Waterloo, And Jesse's tall and blonde. And this thing could not find his face. And the reason is because there were only Japanese engineers in this. And they coded the faces essentially dark eyebrows, and eyes, and so on. So the system, the hand-corder system was extremely brittle. So it's like a system coded to recognize your op.
[7:32] and optimal conditions. but impossible to recognize you're out when running away from a predator. Bye. The machine learning solution is-- Collect a large collection of face images. These will be from all sorts of different types of people. and populations and so on. and then come up with a very simple algorithm that will allow you to detect face. So you can't tell a machine-- How to?
[8:06] You can't write a program that will tell machines exactly what a face is, but you can code a program that will... tell explain to machine these are faces These are not faces. And then let the machine automatically learn what it is that differentiates a face from a non-face. And that's the machine learning approach. It's based on examples as opposed to we telling the rules. We let the machines figure out which are the things, which are the features, we call them features or patterns, or statistics that make a face different than any other image.
[8:47] So the first part is to collect that huge collection of images of faces. You then collect a big data set of known faces And-- The result is this. That's how your face-- detectors are built. Um. I will show you how to do that in the course with a technique called random forests. and so you can have nice consumer electronics and so on. Um. Of course, it's not just for faces. You could do this for anything. any object. So you might take lots of images of pedestrians, and train a machine to recognize pedestrians.
[9:24] Now, why would I want a machine to recognize pedestrians? Thank you. What do you mean? automated cars. Automated cars are not far. They're already approved in California. Um, and, um... Soon. uh... will hopefully have automated cars on the street. Um... I am trying to remember a figure twenty or forty Either 20 or 40.
[9:56] I've got my power set to confuse. Either 20,000 or 40,000 people die every year in the United on car accident. mostly pedestrian. To put this into the importance of this project, in context, imagine the population of UBC dissipated every year. That's the cost. to not driving properly and relying on humans Try it. In India, when I checked a couple years ago when I was visiting, It was 113,000 people.
[10:31] that career which may or may not be bad depending on how you count. Now, um... Actually, it is bad regardless of how you count. 113,000 rises. Huge number of people. Thank you. So-- Same idea. *coughs* You teach a machine, here's a bunch of images, in this case images of cars, And then again, commercial systems these days, like Mobile already are able to detect flames, detect cars and so on.
[11:06] So-- But the key to all these huge innovations, be able to have an automatic-- because without these innovations, it wouldn't be possible at all to have an automatic car. Because how could you deploy a car that can't see a and understand what a pedestrian is in front of it. that would be like murder. Um... the key to it was to come up with this realization that we could teach machines. what they take. So this course, indeed, will be about how to write the programs that let machines grab data And.
[11:42] decide what a face is and and/or what a pedestrian is, and so on. Um. Now, there's other learning machines out there, and those learning machines are humans. Humans are very good. In fact, not just humans, all animals are very good learners. Um. Humans are particularly good. Um. and I'm about to now test. how good humans are at learning with this example. So, I have highlighted here, I have a bunch of images.
[12:15] of objects. which three of them I'm calling tufas. Okay, you're not labeled them by putting a red box around them. So now the question to you guys is how many of you think this one here that I'm pointing with the arrow is a tufa? Okay. the majority. How many of you think this is a two-fly? How about this one? That's a Skeptics.
[12:48] How about this? That a tearful? Just one? So go to an interesting sample, this one here. Thank you. Okay, no, not as many people. Um... This guy here. - This guy? Dear people, Yeah. OK, let's pick someone who likes Let's start with that.
[13:21] Why is that a tufa? God bless you. stem, and it's kind of got roots. and OK, stem, root, and a head. That's what makes it. to do fat. Uh... Who else would like to comment on why they said something is a twofer? or not. - Looking at the dog, the There's a long frown. There is a loose.
[13:52] just sort of Red Bulls. Thank you. Come on. It's hard to describe, but somehow you know. Because you-- very confident to put your hand up or kept your hands down. Even though you couldn't describe it, say that that was a tufa or not a tufa. See? Do you know what a two-finger is? Thank you. most people Anyone else?
[14:23] Thank you. Thank you. So... This illustrates several things. One is how quickly you learn what a two-fay is. You've never seen a two-foot people. maybe in your data field, But you very quickly learn what a 2P is. Thank you. from very few examples. Um... Not all of you have the same opinion on what a truffe is, that's why I asked several questions. So it's somewhat subjective. But there is nonetheless strong agreement Um... And then... There are features that make a twofer be a twofer, whether it's a long stem or the legs head Um...
[14:59] *cough* But by and large, we're not thinking even of those features when we extract them consciously. It's down subconscious. A lot of the feature extraction interpretation of the world is done through the subconscious. Subconscious is very powerful. perhaps more than the conscious. Um. We would like to design machines that are capable of doing this. that can see a few examples of data And just from a few examples, can very quickly learn new concepts.
[15:32] Machines are not capable of making these induction. LEAPS. as Josh and Yvonne calls it. So machine learning is about Being able to extract the features or the patterns or the statistics. The thing that makes us tell what a twofer is, And once you have those features, And you might not do this consciously. We humans don't do this consciously and that's why It's hard. called to try to code things by hand, and instead we have to rely on machines.
[16:07] is to extract the patterns that make a two-fold. The patterns that make a face or a pedestrian And then once you have those, to make a prediction. Where a prediction could be like a decision. deciding whether something is a face or not a face. whether something is a tooth or not a face. or a prediction of a different nature as in, is this thing going to eat me? Should I run away? Is this too dangerous or not? Thank you.
[16:39] There's other types of predictive tasks that are very popular right now, so I'm just going some of them quickly. Um. forecasting. Um... And this I will do as an example in the next class. We're quite often viewing it-- forecast how much energy demand. your building needs, because based on that you can adjust Um... you know, how much energy you buy. And-- you know, potentially safe. Um... Spend the ends of the dollars. for the economy.
[17:10] Thank you. You might want to forecast sales. If you work for Lululemon, you're going to forecast how many people will go for the pink That's a-- at Taskware. Machine learning is actually very useful. Certainly the GAAP and so on stores across North America very-- series of other machine learning. Thank you. You might want--you might want to--you might want to you might have rated a few movies And so you would want a machine to automatically rate the rest of them to protect what you would have how many stars you would have assigned to any other movie so that they can build a better recommendation system.
[17:48] Um... you might want to use machine learning once you teach machine learning what is an HIV virus you want the machine to be able to then detect any departure from this thing that it's monitoring from being a virus. In other words, you want machines to detect mutations. Now those are very important because the moment you detect a mutation you you can choose a different treatment uh... treat the virus.
[18:21] Um. The reason why using your credit cards is so easy these days is because there's very good credit scoring systems. So machine learning classifies that quickly. will determine whether you're-- uh... whether you pass the test or not, whether you're credit worthy or not. So a lot of our economy already-- relies on the usage of very large scale classical Cash in. as well as medical you know, diagnosis.
[18:51] on like noces of mammograms and so on is something that machines eventually do a lot better than um Humans. Ranking, another task where machine learning is very useful. So if you use Google image search for example, quite often you do a search. and then you click on an image. When you do that, you're basically telling Google, Hey Google, for this word, This is the image that matches it.
[19:21] And this is how Google Image Search got better and better over the time because we were--we All humans were labeling the images by clicking. Often you do a search for a machine learning term. And at the top, the top 50 images shown, there's these pictures of hot models. And that's because that's where people click when they search for support vector machines or whatever technique they want. So there's some failures to the algorithm, which is what makes what made me realize what it is that they were doing.
[19:54] Thank you. uh... And of course-- Um... In this huge flow of information that we live in today, being able to summarize information is also important. And already there's lots of tools that will try to condense information for you, whether it's news or any other type of information and make it available. One of these tools is-- um, was Zite, and that was built by two students who take took this course many years ago, my class and Eric Berchu.
[20:26] They, together with some venture-- capitals here in BC eventually built this company that was sold to CNN. If you go to CNN website, the algorithms that power CNN were developed here. Okay, so... and finally Um. Often talking about AI, I think machine learning is the real AI. certainly a huge component of AI.
[20:58] or there is needed in order to build intelligent machines, but it's certainly a huge component And great progress has been made in the last decade. Thanks so much, Imran. Thank you. Now, when do we need machine learning? When our expertise is absent. If you want to send a robot to Mars, Um... Remote control is not going to work, because it takes too long to transmit the signal. So you need a robot that can understand the environment, and it can repair itself, and so on. Thank you. Um... From the video, it was clear that we can't understand how we see or recognize.
[21:34] of things, or even we don't even understand how We-- see concepts from sounds. And so-- lack of ability to introspect motivates the use of learning Um... It's also useful when you're dealing with problems that keep changing. Like if you're trying to give someone good recommendations for, I don't know, kitties or dogs or whatever it is that people recommend these days on the web.
[22:06] Um. People's interest change with time. And so if you relied on curators, Um... that would be--would require a lot of courage. So it's important to Um. to automate to have automatic systems so that As people's interests change with time, you can still provide them with the content they need. Um. Connected to that is-- this problem that often everyone is different. If I wanted to recommend news, I wouldn't have to build a news recommendation system for each different person.
[22:41] because each person That means that if I have a million customers, I will need, say, to the-- instead of having a million curators, what I want to try to have is a million automatic recommending contents. Um... And of course-- Um. for problems where we humans can't reason. And when you work for a search engine, for a company doing web analytics, then it soon becomes clear that you can't rely on rules of-- you can't rely on people to do things The scale of the problem is so large, you do really have to of relying on machine learning.
[23:31] Now, data. Data has-- increased a lot over the last few years. In Um, as-- Lon, Halaby, Peter Norbeck, and Fernando Pereira, researchers at Google, from some of the leading researchers on language at Google. Um. point out in the paper. in 2009. In 1960, seven Um. One million words was considered a large data set. In 2006, I don't even know what that number is.
[24:03] Um-- Trillion. That's it. Thank you. uh... Thank you. Word dataset is what was regarded a large dataset set in academia. So there's been huge orders of magnitude. on the data sets that you guys are capable of actually managing in your personal conversations. And then if you go to Amazon EC2 and so on, it gets even more fun. What were the success stories of language.
[24:37] I'm out. over the last few years. technological success. Watson. Pardon? What's that? Watson? Jeffrey Lee Chan. Cute. success story. Success stories as-- And that they've actually created revenue for a lot of people and solved people's problems. Google Translate. ERIC SCHMIDT: But Watson is a good one. Google Translate. That's a brilliant one. Because... Um.
[25:07] My wife can be at Aritzia doing her job, and she needs textiles from India, and she needs someone to weave those textiles in China, And she's sending emails to all these people in India China. who might be speaking Hindi, who might be speaking Mandarin, and they're all able to communicate because they're all using Google Translate. So it completely revolutionizes e-commerce. because now the world becomes a smaller place. You don't even need to all speak the same language in order to communicate.
[25:40] Occasionally, there's some problems with that. That's pretty much how people do things. even in small companies in Vancouver, this is how--this is how--this is how we're going corporations are able to engaging global business. with their product. What's another example? Apple marketed it. heavily. suit Speech recognition. So, it's picture recognition theory and So-so. Many particles.
[26:12] We're in Canada. But-- Googles. speech recognition works extremely well. I don't know if you've tried it recently. If speech recognition on mobile phone. The beautiful thing as well is that every time you get it right or wrong, You provide feedback. for Google. get better. So I'm assuming that speech recognition is one of those solved. Um. How-- you Thank you.
[26:45] combined teacher of the year Cheers. It's very strange. by my books. Pardon? Something like real time translation. into another language. And... Oh, yeah. You can even do both of those. So eventually we'll be able to speak if you like. Yeah, like science fiction. Where we can all... Maybe I'll tap talk. Father. I wouldn't be surprised.
[27:17] In terms of speech recognition using deep nets, Um. something I'm going to come to later in the course. There's two teams doing it. The first team that started it was actually in soft. And then Google got started on it. And so they're using slightly the same-- sort of infrastructure. about that. common thing among all these? is that there's lots of data. There's lots of data in two languages. And there's-- lots of-- in other words, match data. So this is the text in French and this is the translation.
[27:55] um And the same for speech recognition. People created this huge data. sets. Moreover, the data sets keep getting larger by the minute because where companies like Google and Microsoft and so on have been brilliant is that they give you a service and if you don't like the translation They allow you to correct it. And now because of that, you're always providing data for these companies to get better. at providing their products. So they're harnessing the crowds.
[28:26] Thank you. uh... Now, for a lot of these applications, you do need a lot of data. A lot of data is essential. If you're going to do speech recognition, Really, the stumbling block is not even knowing how to build the technology. Um... but it's actually having the data. to train technology. Data is the commodity. Yeah. Here's an example to illustrate this point. by Alyosha Efers at CMU. So he had initially an image that he didn't quite like because it had a house So what he wanted was-- could I remove this house?
[29:05] So you can obviously remove it. And then the question is, If you have a collection of other images from the web, Can you use this other collection to fill this in? You know. in a way that will look better than this. If-- he found that if he used a few images, it didn't work. But then when he increased this, She was able to have the machines automatically composing new images That's-- still look very plausible. For this to happen, you really need a lot of data.
[29:42] Now, coming back to the Google guys, they say, In their paper-- that they've solved the problem of how to get people to contribute data. And-- Certainly Facebook and Google have done a great job at this. Um... Dave? They claim to have solved the technological problem of aggregating and indexing the data. They have done this to some extent, and Google certainly has done a good job. There's still a lot to be done.
[30:14] But most importantly is the fact that We still have not interpreted the content. So Here's an example, citation matching. Now, when you do a search in Google, you essentially use an inverted index and you look for that term in the search and you check if that exact word appears in the web page and then you bring back those web pages. And then there's some ranking that gets done using an algorithm which--if you do the Python exercise in the Python website, that's the--that will be a way of learning to learn how to code a Google search engine.
[31:04] Now... However, the problem with doing search this way is it's very much based on popularity. Google is a popularity machine. Now, a counter example is citation matching. So citation matching is the problem of trying to decide whether two citations. So in scientific papers, you always cite other people's work at the end. So this is something you will do when you write your report for your project.
[31:38] Deciding whether two citations are the same is actually proved to be extremely hard. And that's because people often misspell names. Often people write the volume or not, include the pages, or have the wrong pages. And what tends to happen a lot with academics is if I cite a work and I didn't get the page number right, And then someone else. wants to cite my work, but my work really goes back to the other paper. They haven't read the other paper, but they just cite it nonetheless. And what they do is they copy the citation.
[32:14] And also you're just a couple of hours behind the deadline, you have to submit a paper that's just cut and paste with a bunch of citations. You'll see what I mean when it comes to handing the project at the end. And so the problem with this is that if something is wrong, it can very easily propagate. Now you have 898 citations with the wrong page number and 22 with the right page number. If you just go with popularity, you will pick those. So you have to do something else. You have to actually do an intervention to be able to solve that problem.
[32:54] Let me show you another example of this. um This is the issue of information extraction. And I'd love to see a few projects on this topic in the course. So this information extraction engine is by Oren Ezioni and Dan Weld, who are professors at the University of Washington in Seattle.
[33:26] And so this allows you to type things like Obama. Married. And now we're starting to do something a bit more interesting, something with semantic value, not just a query, like, you know, not just-- putting in like Obama married and retrieving a bunch of pages that say Obama married but you actually want to know who married Obama who or who did he marry and if you do a search And they do this by extracting triplets of noun, verb, noun from the web.
[34:08] And so that is what you get and the numbers I'm imagining is the frequency. That's the web. The web has truth, but the web is also full of garbage. And extracting information reliably from the web just by using counts popularity is extremely hard. And the only way you can fix this is by intervening. And well, there's two ways to go about it. Right now, they use a bunch of rules. And this is the best, by the way.
[34:44] This is the cream, what you're looking at here, as to what we can do today. This is the sort of thing that Google will use. and so on. likely and a lot of companies use it. Orin Asuncion is a great entrepreneur and he's sold many companies using this type of technology. If we are to go beyond this, there are two routes. One is to start parameterizing these rules. Because right now, linguists tend to create all these rules of syntax and so on.
[35:19] It has to be a noun phrase followed by a verb phrase followed by a noun phrase and so on. Or then they start adding a bunch of exceptions. If you could just parameterize all these and then having a learning machine estimate what are the best, predict labels, and then have people vote yay/nay on labels. If you could create a cool app that allows people to do that, we would learn to do information extraction. a lot better. But if you want to uncover truth, ultimately we need to contend with the concept of causality. We need to actually know what is... we need to conduct an experiment.
[35:56] or we need to actually probe, we need to conduct an action to be able to actually figure out. We would have to do something now, after this point. We would have to click, we would have to do some research, and we would have to decide how we're going to do that research in order to be able to discover truth. So action is necessary for truth. That brings me to the next part.
[36:38] which is-- Learning is not something that just happens where you have a lot of data and then you learn from it. That's how I did it in 340, for those there. But in 540, we're also going to look at when we actually actively go and seek the data, when we do interventions. Because most systems are not just about... being flooded with data and doing something as a result, but you actually do interventions in the world. And most of the data you know, in order to get data we conduct experiments. I mean, this is pretty much at the heart of the scientific method.
[37:20] If we want to understand something we need to perform actions. So, I'm going to do a bit of causality this year in this course because I think it's important. eventually to have true understanding. Okay. The course will also have a lot of neural networks. And recently, there's been a lot of cool stuff happening in an area called deep learning. How many of you saw the Google Frontline deep learning network in the New York Times?
[37:59] One, so machine learning was big in the news. Big enough that two people, three people, four, five managed to see it. It was in the front page of the New York Times. So Google was claiming that they had built this neural network that can look at YouTube videos and just by doing that it can learn what a cat is or what a face is and so on. And so we're going to go carefully over that experiment in this course.
[38:29] state of the art in deep learning. But needless to say, there's been some really amazing advances over the last three years in this area that are bringing us much closer to doing active general object recognition. The speech recognition that got a lot better over the last--at Microsoft and Google, the reason why it got better was because of this deep learning. And so quite a bit of the course, and it's all very biologically inspired. And so we are going to cover that in the course.
[39:03] It's going to be almost like a third of the course. I'm not going to go into neuroscience, but I will not be on this little caricature here on the screen, which is there's a brain. If you're looking at vision information, there's a visual signal that comes into your eye. Again, that visual signal is very tiny, and it's dynamic, and it's about the size of your, the high-resolution bit is the size of your thumb.
[39:37] - At this distance. and then the less, the lower resolution about the size of your hand. But you really don't see much. This is really mostly imagined. Then there is--everything goes by-- this optic nerve to the back of your brain which is the first part of vision and try doing this, try hitting the back of your head. Things go black. There's a reason. That's where your vision is.
[40:10] Over here with LGNAs there's about a million fibers. So that starts giving a rough idea of what the resolution of that image is, thousand by a thousand. And then at the back there, there's an area called V1, which is very interesting. Now I'm going to have to say a few things about that. Now the brain is a good memory device and it's a predictive device. The way we store things in the brain is in an associative manner.
[40:45] And what I mean by that is, for example, something like that, where there's--if you want to retrieve something, when you present a new query, like in this case, a plane with clouds, you retrieve a plane. So it understands that there's a plane there. um Who doesn't mind giving their phone away? Can you say your phone number loud forward? 604-442-7185.
[41:19] Can you say it backwards? I heard 280. - Okay, so significantly slower when he says it backward. And the reason is because you don't really process numbers like a computer does. You don't store numbers, like the, you know, with a random access memory or whatever. What you've learned is this pattern that you repeat, which is 7782391190. So easy, 'cause it's like this, you store that pattern, which is your phone number. You haven't stored the numbers.
[41:55] stored a pattern that sequence same is true with the alphabet a lot of people can have no problem saying the alphabet forward but people struggle saying the alphabet backward because the what they've learned stored is patterns okay sequences and so on so i have a different experience so i'm not a native english speaker but when i'm telling somebody my phone number using chinese pretty fast. So when I'm speaking it using English it takes some time.
[42:33] So my feeling is when I'm speaking my phone number using English, I... the picture of my phone number and read that picture. And I think that might be the part of where time is . Yeah, it's true. But you're doing it by association. So association, by the way, here I did association with time, but it could be spatial. In fact, a lot of people who are very good at memorizing things, it's because they think of an image.
[43:07] Like if I give you a list of 100 objects, then what they try to do, 100 is actually an overkill, 10 objects. So most people would find it really hard to remember 10 objects. But if I, the ones that are good at it, they compose an image with the 10 objects. and then they just recite that image. But they create a pattern and a context, and that makes it easier. somewhat similar to what you're saying. The key point being that memory is associated.
[43:41] one game for today. If you were to pick one of these to be a spider Which one would it be? A jumping spider. How many of you have seen a jumping spider? They're all over the place. So how many people would vote for this first guy being a jumping spider? You only get to pick one, so there's two votes for this one. This one. Zero. Next.
[44:12] Zero. Next. Three volts, four. Next. Four. Next. One, three. This guy. One vote, two, three. Next guy. Okay, so that's like what about 30 Okay. Now...
[44:45] If we compare these numbers, so those were images that were shown to a male jumping spider to see if that male jumping spider would see this as a female that wants to have spider sex. You never expected I was going to pull this one on you. So-- and The point here that I want to illustrate is that even little animals are not that far from us in terms of their ability to recognize us.
[45:15] So spiders are actually able to recognize simple caricatures, drawings. So not just actual images of-- that's a jumping spider. And here you see those jumping spider recognizing can recognize mating behaviors and so on, even when it just of caricatures and so on. So spiders Also, these jumping spiders, they're able to recognize 3D depth. and so on, which is essential to travel in the world. So, we don't need to get to the high-level brains.
[45:48] I started with the high level brains, they're showing a human brain. but the very low level brain of the spider is an extremely sophisticated piece of machinery that We still are struggling to replicate with our own electronic devices. And evolution is beautiful, just like spiders have learned all these amazing-- uh... mental traits, recognition and so on. There's creatures that have evolved, like this moth, that they've evolved to replicate to look like jumping spiders so that they Don't fall prey to them.
[46:26] Um... Most of what we see in the world gets encoded in our brain, but even more, the location of things in the world. Like in this case, what we see in red. So that's a box. And that's square. And there's a raft that moves in the box. And then what we're monitoring is whether a neuron fires or not. And as you can see, there are particular portions periodic interval in the box. where a neuron fires. And that's how you can tell where you are. close your eyes just let that neuron flow.
[46:58] fire tells me where you are. Where you are is encoded in your brain. Now, going back to vision, back to this, All these fibers get into these first level of cells. and V1. And people did the same sort of experiments. there probes. on cats. Thank you. at the back of the brain And then they shortcuts these bars as shown in the screen.
[47:30] vertical bars and horizontal bars. And when you do that, Um... The experiment is by Hubel and Wiesel. You can actually see the whole video in YouTube, if you Google Hubel and Wiesel. Um-- What they found was the following. If-- If the bar is vertical-- is horizontal, sorry-- as shown here in the first drawing. If the bar is horizontal-- that the neuron doesn't fire. The neuron only fires when the bar is at a particular angle.
[48:05] Angle is where the neuron starts firing. That means that each of these neurons is basically tuned to-- vertical, specific lines. Lines are different. orientation. So all these neurons basically have an age map. of the world. because they fire-- different neurons will fire for different edges. Now there's many neurons. there in V1. Each of the little boxes in this collage is one of those neurons. And each of those will-- all 20 million or so.
[48:37] will be monitoring checking whether there's a patch here, whether there's coordinate there. whether there's an edge here. And it's all those features Not just the eyebrows. But it's these 20 million features, when you take compositions of them, that allow us to recognize whether something's a face or not a face. They're all very basic. And it's-- But in having these huge compositions, we're able to do all sorts of recognition. Thank you.
[49:08] So the basic idea-- and we will go over this in the course I'm just... At this stage, I'm just giving an overview which hopefully will inspire you. to continue. Um. And later in the course, we'll actually look at the details of how we build these. uh... But the basic idea is that you will have neurons And each neuron is tuned to one particular stimuli. Like in this case, an edge. that happens in this part of the image. Thank you. So our nature is basically a contrast between light and dark.
[49:41] And when that neuron sees this, it will fire. FIKE gets produced. several spikes start getting produced. If a neuron doesn't see this, it doesn't fight. And this is the basis for constructing a neural network. something that looks at an image and then when you have 2000 neurons connected to the entire image and some of this units will fire, some of them will not fire. The implication of this is that we're mapping an image to a code.
[50:17] of about 2,000 years. We will see in this course how to learn these. how to learn, what are the-- receptive fields for each neuron. what's What should each neuron be paying attention to? What ages? So this we will learn. for each of these neurons We will learn them automatically from data. And then what researchers have done and deep learning is that they then take the output of these neurons and they feed them to a second network.
[50:50] where they learn again the parameters of that network. And they find that they learn things like the subject parts, and so on, when they use images. when they just train it with, say, images of faces. And then further down up here, you have actually entire faces. So, all this gets done automatically these days and this is essentially what Google was Google was taking one of these large networks and was able to just train it And-- And the way it's trained here will be without supervision.
[51:27] So that's going to be an interesting We're not going to be saying these are images of faces are not images of faces. We will learn how to take a huge collection of images And just by using some concepts, which is basically what you imagine should look like what you see. that very basic idea. If I have a learning system that's capable of imagining Um-- what it sees. it should be able to learn interesting concepts. And that's essentially what Google demonstrated. They took a system where you took many images, The idea had been demonstrated to be fair by many groups in academia.
[52:06] But Google just took it to the next level, very large scale. And so they-- by just training on all sorts of images in YouTube neurons that were capable of recognizing faces emerged. So there's one particular neuron in this ensemble of neurons Every time it sees one of these faces, it fires. So now that's also biologically plausible. There's-- paper. There's a very famous story of the Halle Berry Neuron where neuroscience actually conducted a study where you know, Which animal did I use?
[52:42] for that from the chat. Probably a cat. uh... Humans are not allowed in these things. Um-- that that fires-- has a neuron that fires every time this cat sees Halle Berry, a picture of Halle Berry. Thank you. Thank you. and uh... Thank you. And then they can also, in practice, process. switch this neuron on and because you've learned the receptive fields, you can actually learn, "What is the input that makes this neuron fire?"
[53:18] And that was the picture that made it to the front of the New York Times. century. We're going to see how to compute this picture in the rest of the course. I will also see as to how The same idea that you can reconstruct the input can be used in order to do this sort of-- tracking and be able to complete this seen. in order to make this open. protections. Um... Finally, to end this, this is... the speech recognition that works now these days on your phone It does a first level which is it goes from sounds come to central coefficients. That's sort of very easy.
[53:59] - Two, two? Standard. It then uses a deep learning, again, a neural network to go from those basically sound features. to features that can be fed into something called an HMM. in order to do recognition. So throughout the course, we'll go and look at the elements Um-- that are necessary sorry to publish And that's pretty much it. So the next class will start already with--with the--with the covering supervised learning.
[54:30] And we'll start with linear models because they happen to be very easy to understand and I'll let him hold it. Thank you. Bye. Thank you. Thank you. Next time, I want Russia.
[55:10] Yes, yes. The station is a large, multi-use station with a capacity of about 1,000 people. The station is a large, multi-use station with a capacity of about 1,000 people.
[56:10] Thank you. Yeah, I know.
[57:10] Thank you. *Captions by Project readOn* Thank you.
[58:40] I'm just sitting there. I don't think we're in a federal situation.
[59:40] .
在Vidleaf工作台中打开
搜索转录文本，选择行，复制带时间戳的引文，翻译。
署名
Nando de Freitas 的《机器学习——导论》（https://www.youtube.com/@profnandodf），依据 CC BY 3.0 许可（https://creativecommons.org/licenses/by/3.0/）。源视频：https://www.youtube.com/watch?v=w2OtwL5T1ow。本页是该视频的文本转录，添加了段落分隔和时间戳；创作者与 Vidleaf 无关联，也不为其背书。Vidleaf 添加了 AI 标点和 AI 生成的摘要；原始措辞仍可在本页查看。
您是创作者或权利持有人吗？请求更正或移除：copyright@vidleaf.app（参见关于这些页面）。
最后更新于