2019原文書閱讀筆記：How Smart Machines Think

 

How Smart Machines Think

作者：Sean Gerrish
出版社：The MIT Press
原文精裝書October 2018出版
原文平裝書October 2019出版
312頁
出版社官網書介：https://mitpress.mit.edu/books/how-smart-machines-think

這本書本來是在版權評估書單之中的。但經過內部審閱和討論之後，感覺「口味」比較偏向天XXX、XX周X會出版的書籍，所以最後就沒有納入考慮。但在審閱期間我卻讀出了興趣，於是利用了某段書與書製作期間的空檔，把整本書稿都完食了。閱讀期間真心覺得這本書非常好看，時不時甚至會有「不能親手編譯這本書真的好可惜啊」的感慨（但這就不是我能夠決定的咩），只能期待將來會有中文版囉，不然我還真想買一本精裝版來收藏呢（我應該要寫信給天XXX、XX周X，推薦他們考慮出版這本書的中譯本才是！）

本書的目標讀者，就是像我一樣的「非理科生」：「對機器學習／人工智慧／深度學習」等題目有興趣，略懂皮毛、似懂非懂，卻又希望找到「更淺顯易懂的敘述」來獲得知識。如果你跟我一樣處於這種「不上不下」的階段，卻又想要進修一些新知，無論是為了理解最近這波AI熱潮、為了提升職場競爭力、或是想要增進英文閱讀能力，甚至只是單純想要閱讀一本好書…無論是何種情況，我都非常推薦閱讀這本原文書！

這本書一共17個章節，但可以簡略區分為五大討論主題：

CH01-04：車輛的自動駕駛

Ch05-06：Netflix的影片推薦

CH07-10：詳細解說artificial neural networks是什麼、如何運作，以及如何「玩」Atari video games

CH11-12：簡介自然語言處理（Natural Language Processing，NLP）及其神奇之處；IBM的Watson如何在Jeopardy!這樣的益智競賽之中大放異彩

CH13-17：其他的應用，主要還是遊戲方面居多，例如：圍棋（GO）；還有對未來（五十年後）AI人工智慧的發展及其應用的預測

以上章節內容，（從一個非科技領域讀者的角度來看），我覺得最精采的是描述「自動駕駛」和「影片推薦系統」的這六個章節（CH01-06）。或許是因為這兩大部分都有提到「古早的歷史」，更進一步講解了各個團隊之間「彼此競爭又互相學習的競賽過程」吧！我喜歡閱讀那些「不是特別一帆風順」的開發過程，看開發人員和設計人員如何從失敗之中學習或修正，看「雛形」或「原型」如何一步一腳印地成為當今炙手可熱的「商業應用」或「熱門話題」…等。雖然有些已成歷史，有些尚在演進，但這些前人的寶貴經歷，卻扎實地替人工智慧的發展及其實踐打下了豐厚的基礎，非常值得一讀。

以下（落落長的紀錄），是閱讀本書時，英文佳句或英文句型的摘錄，學習（新）單字的紀錄，以及一些心得感想。分享給有興趣的讀者參考。

國鳳　2019/06/27

ＸＸＸＸＸＸＸＸ👉我是分隔線👈ＸＸＸＸＸＸＸＸ

P.ix

Arcane = mysterious; hidden; secret

P.x

How Smart Machines Think assumes relatively little background in machine learning or computer science, and hence is entirely approachable to a broader audience.

👉作者在Preface說明這本書適合什麼樣的讀者，或者需要什麼樣的知識背景才能充分理解。作者這段話的意思是，討論機器學習或電腦科學這兩大領域／主題時，這本書假設讀者需要「相對較少的相關知識背景」，因此讀起來更平易近人、受眾也更廣泛。

Behooves = appropriate or suitable

P.xii

Eventually I realized that I should share what I was learning during my own research with other people, so they wouldn’t need to jump through the same hoops to understand the same things. In other words: I wrote this book because it was a book I wanted to read.

My hope is that there is something new in this book for everyone.

P.4

Pendulum鐘擺

P.6

Defecate排泄物；排便

I’ve written this book for anyone interested in how these devices work. You won’t need to have a college degree in computer science to understand this book, although I’ll assume that you’re familiar with some basic facts about computers, such as that they follow explicit instructions encoded by humans, that images are represented by computers based on the amount of red, green, and blue they have in each pixel, and so on.

👉作者再次重申，讀者不一定要有電腦科學或是IT相關背景才能理解本書；這本書是寫給「任何有興趣知道這些裝置如何運作的人」。但讀者對電腦是如何運作等方面必須要有一定基礎的了解，例如，電腦是依循人類編碼的明確指令來運作的，而電腦所呈現的圖像，則是以每個pixel當中的紅色量、綠色量和藍色量為基礎，比例不同，呈現的顏色就不同，等等。

And finally, I’ve written this book so that you can usually jump straight to the topic that most interests you if you don’t feel like reading all the way through.

👉如果不想閱讀整本書的話，作者表示他寫的這本書能讓你直接跳到你最感興趣的題目，可以依照自己有興趣的主題來跳著閱讀，不一定得按照章節順序來閱讀。

P.7

In the first half of the book, I’ll outline some of the key ideas that enable intelligent machines to perceive and interact with the world.

👉本書前半段的內容概述

Later in this book we’ll look more closely at how computers can play a variety of games.

👉本書後半段的內容概述

P.9

Burrow （動物挖掘居住的）洞穴或地道

P.14

Augment = increase

P.19

Rudimentary = elementary

P.21

Debacle = a complete failure

P.22

DARPA officials were also excited, congratulating each other about the race. For the past eight years, the field of self-driving cars had been hibernating in virtual winter ever since Ernst Dickmanns, one of its leaders, had proclaimed that the field would need to wait until computers were more powerful.

P.24

The winner of the race was Stanley, a car built by the Stanford Racing Team, newcomers to the race that year. Stanley drove so fast that it had to be paused twice to give the car in front of it more time. Eventually race organizers paused the car ahead of Stanley to let it pass. In the end, Stanley finished over tem minutes faster than the Red Team’s Humvee.

👉知名的Stanford自駕車在第3章登場了！個人觀點，第3章講述開發Stanley的部分，堪稱本書最值得一讀的一章！

The Stanford Racing Team had the insight that placing so much emphasis on mapping and navigation at the expense of sensing the environment was a mistake.

👉at the expense of sth的意思、解釋及翻譯：If you do one thing at the expense of another, doing the first thing harms the second thing；在犧牲（或損害）…的情况下；以…為代價

P.28

Veer 改變方向；轉向

P.30

Heuristically 啟發式的、探索式的

（根據Cambridge Dictionary的英英解釋， heuristics是：(of a method of teaching) allowing students to learn by discovering things themselves and learning from their own experiences rather than by telling them things）

P.31

Switchback 之字形的路、Z字形的山路

P.31-32

To understand how Stanley did this, imagine you’re a vampire who has just done a load of laundry. Since you’re a vampire, your favorite colors are red and black, and your socks are various shades of red and various shades of gray. After coming home from the Laundromat, you begin to sort through these socks, spreading them out on the bed so that similarly colored socks are near each other. Over time there will be a pile of red socks and a pile of gray socks, and they might overlap where the darker shades of red meet the darker shades of gray.

👉不知道（在機器學習當中）所謂的分類器classifier和所謂的集群／叢集clustering是啥咪碗糕，作者提供了這段非常有畫面的比喻，我覺得很不錯！粗略的翻譯是：請讀者想像自己是一個剛洗完衣服的吸血鬼。因爲你是一個吸血鬼，你最喜歡的顏色是紅色和黑色，你的襪子是各種紅色的陰影和各種灰色的陰影。開始排列／分類這些襪子時，顏色相近的襪子就擺在彼此附近。然後，你將會有一堆紅色的襪子和一堆灰色的襪子，它們可能會在「顏色較深的紅色」與「顏色較深的灰色」之處，「交會」「重疊」。

But then imagine that you find a bright green sock in your laundry. This sock clearly doesn’t belong in either of these piles, so you conclude that it must have gotten mixed up with your clothes at the laundromat. You reject it.

👉簡略的翻譯是：但是想像一下，你發現了一隻亮綠色的襪子。這隻襪子很明顯地不屬於這堆襪子中的任何一堆，所以結論是，它一定是在洗衣店時，和你的衣服混在一起了。 您（要）拒絕它／捨棄它。

P.34

Swerve 突然轉向、轉彎

Erratically　不定地、怪異地

P.57

As robotics departments were busy in 2006 preparing their cars for the DARPA Urban Challenge the following year, Netflix made an announcement to the budding data science community about their own Grand Prize: they were looking for teams that could create movie recommendation engines, and they were willing to pay $1 million to the best team.

When Netflix made their announcement, their streaming video business didn’t yet exist; the company operated as a physical DVD rental service. Customers could request DVDs from Netflix, and Netflix would send these DVDs to them by mail. But customers needed to give up one of their current DVDs to receive the next one, and the new DVD might take days to reach them. A bad selection could ruin days of quality movie-watching time, so customers tended to be careful about how they made their requests. This is where Netflix’s desire to recommend movies came in.

👉大略的翻譯是：

2006年，當機器人部門正忙着準備自己的自動駕駛汽車，好參加隔年的DARPA Urban Challenge競賽時，Netflix向資料科學團體們宣佈了自己的大獎：他們正在尋找能創建電影推薦引擎的團隊，且他們願意向最好的團隊支付100萬美元。

當Netflix做出宣佈時，他們的影音串流業務尚未存在；該公司的主要營運只是一個實體DVD的租賃服務。客戶可以預約申請Netflix的DVD，Netflix將透過郵件將這些DVD寄送給他們。但是，客戶需要寄回手中擁有的其中一張DVD，以換取收受下一張DVD，而新DVD可能需要數天才能送達。糟糕的選擇可能會毀掉寶貴的電影觀看時間，因此客戶往往會謹慎地考慮他們該預約哪些電影。這就是Netflix想要建立電影推薦系統的原因。

哇！2006年的100萬美元耶！對照Netflix今日的成就，它肯投資、肯花大錢辦活動吸引人才，也算是成功的要素之一吧？而關於郵寄和預約實體DVD的描述，也勾起不少我在美國求學的回憶：百視達是在我升大二時關門大吉的，而實體DVD出租店／出租服務的式微，和近十年線上影音串流服務的蓬勃發展，亦形成（既心酸又）強烈的對比。

P.60

Imagine that you’re editing a cookbook called The World’s Best Recipes

for Kids. For this cookbook, you’ll collect recipes that appear on the website

Bettycrocker.com. You have a simple decision to make for each recipe: is it, or is it not, a recipe you should include in the kids’ cookbook? One way to answer this question would be to prepare each recipe you found on the website, feed it to your kids, and ask them for their opinion.

But if there were 15,000 recipes on this website, then even at a healthy clip of 9 new recipes per day, you’d be cooking for over four years. How could you determine which recipes are good for kids without a huge investment of time and energy?

A machine-learning student would eagerly tell you how to solve the problem: you could train a classifier! In the field of machine learning, a classifier provides a way to automatically figure out whether an item (like a recipe) belongs to a certain category—like “recipes that are appropriate for kids,” as opposed to “recipes that are inappropriate for kids.”

👉這一小節的標題是HOW TO TRAIN A CLASSIFIER，如何訓練一個分類器呢？以下是我對這段（有趣）敘述的粗略翻譯：

想像一下，你正在編輯一本叫做「世上最好的兒童食譜」的烹飪書籍。您要為這本食譜收集那些出現在Bettycrocker.com網站上的食譜。每個食譜都有一個簡單的決策：它「是」、或者「不是」一道「你應該包含在兒童食譜當中」的一道菜色？

有一個能夠回答這個問題的方法，就是親自料理你在網站上找到的每一道食譜，把它餵給你的孩子們吃，然後詢問他們的意見。但是如果這個網站上有一萬五千個食譜，那麼即使每天有九個新食譜的健康時段，你也必須花費四年以上的時間來烹飪。如果沒有投資大量的時間和精力，你怎麼能夠確定哪些食譜對孩子有益呢？

一個機器學習的學生會急切地告訴你如何解決這個問題：你可以訓練一個分類器！在機器學習領域，分類器提供了一種方法，可以自動確定某個項目（例如：食譜）否屬於「適合兒童」的特定類別食譜，而不是「不適合兒童的食譜」。

以上諸如此類「生活化又淺顯易懂」的比喻，是本書解說人工智慧、機器學習、深度學習和強化學習時的優點和特色。

P.98

Temporal discounting = time adjustment = 時間折價

P.100

Trial and error = 試錯；嘗試錯誤

Nudge 輕推、推力

P.101

temporal difference, or TD, learning 時間差分學習（演算法）

P.102

But once we know the agent’s current state, we can forget about everything before that, because we assume its current state captures all of the history that’s relevant in anticipating its future. This is often called a Markovian assumption.

P.109

Fraught充滿了…；伴隨…

P.111

A neural network is a computer, and it is therefore a prime building block for an automaton.

P.127

The Turk 土耳其行棋傀儡

P.128

Ruse = a trick

P.129

…we’ll spend the rest of this chapter emphatically digging more deeply into some of the details behind how artificial neural networks—particularly deep neural networks—work; and we’ll start by creating a neural network that can recognize photos of dogs.

👉CH09的大綱概述

Propagate = to produce

P.130

The computer does all of the hard work for us, and we just need to feed the network as many training examples as we can find for it. If we were fitting a network to classify images, we would repeat this process with image after image, and we’d repeat the process until the network was no longer improving. As long as we have enough data and a big enough network, we could train the neural network to recognize just about anything we want it to recognize.

👉簡言之，資料「餵」得越多、越豐富，成果就越好。

P.131

Squiggles 彎彎曲曲的線條

Overfitting can become problematic because it might make assumptions about the data—assumptions like “lots of green in a photo means there isn’t a dog in it”—when it’s not justified in making these assumptions.

👉這一小節講解了什麼是Overfitting，也就是過適、過度擬合、過度學習、過度訓練的意思。過度學習（被餵食的）訓練資料，變得無法順利預測「其他」不是在訓練資料之中的資料。

Remiss 疏忽

P.133

If you don’t have lots of photos to train your network to find pictures of your dog, then you will very possibly overfit the neural network……We’ll explore both of those now, starting with having lots of data.

👉造成Overfitting的原因之一就是訓練資料太少了，所以要取得更多資料。

P.148

Dilute 稀釋

P.155

Correlation 相關

P.159

recurrent neural network = RNN = 遞歸神經網路、 循環神經網路

The only difference is that convolutional filters that share the same weights don’t typically feed into one another. The very nature of RNNs, on the other hand, is that each RNN unit feeds its output directly into the next RNN unit,…

P.168

ADVERSARIAL 對立的；對抗的；敵對的

generative adversarial networks (GANS)  GANs生成對抗網路

P.173

Coleslaw 美式的高麗菜沙拉

P.176-177

They were crowding around televisions at the bar, three people deep, to watch Ken Jennings during his famous winning streak.

👉winning streak是片語，有「接連獲勝」、「連贏」、「連勝紀錄」的意思。

But in the end, inspired by the possibility of success and some hunches about how they might proceed, they relented, and Watson was born.

👉Hunches = 直覺；relented = 緩和

Scatterplot 散佈圖、散點圖

But their converted system performed abysmally…

👉abysmally = 極糟糕地；可怕地

P.178

…analyze the question, come up with candidate answers with search engines, research these answers, and score these answers based on the evidence it found for them…

👉DeepQA運作流程簡述

Natural language processing, or NLP 自然語言處理

P.179

The most important task for Watson during its Question Analysis phase was to find the phrase in the clue summarizing what specifically it is asking for. Take this clue, for example:

It’s the B form of this inflammation of the liver that’s spread by some kinds of personal contact.

The phrase summarizing what the clue is asking for is this inflammation of the liver. Watson’s researchers called this phrase the “focus.” The focus is the part of the clue that, if replaced by the answer, turns the clue into a statement of fact.  If we replace the focus of the clue above by the answer, hepatitis, it becomes:

It’s the B form of hepatitis that’s spread by some kinds of personal contact.

👉我覺得這裡的focus有點接近我們中文所說的「文眼」的意思，也就是「核心」「關鍵」的字眼。只要Watson能辨認出／抓出this inflammation of the liver這句「文眼」，替換成hepatitis，就是事實／答案了。

P.181

Parsing分析

Parse tree 分析數

P.183

In this clue, the ambiguity is around whether it’s the inflammation that’s spread by some kinds of personal contact, or whether it’s the liver that’s spread by some kinds of personal contact. While it’s painfully obvious to us humans that livers can’t spread by personal contact, this isn’t obvious to Watson’s sentence parser. There’s nothing ungrammatical about that parse, even if it’s semantically weird.

👉這裡作者的討論延續了剛才肝炎（hepatitis）的例子。有時候Watson會遇到類似這種「語意模糊不清」或因為句型結構而產生「歧義」的情況。究竟是「透過某種人與人之間的接觸所傳播的inflammation」還是「透過某種人與人之間的接觸所傳播的liver」？對我們真人來說，liver，也就是肝臟，是無法透過人與人之間的接觸來「傳播」的，能傳播的只有疾病、病症或是發炎的徵兆，所以很明顯地，答案必須是「究竟是哪一種inflammation呢？」（而不是「究竟是哪一種肝臟呢？」）但因為語意的歧義，對我們真人來說很明顯的「常識」，對Watson而言，卻無法清晰地辨認；讀到這裡，感覺Watson面對的困難，就跟我們亞洲人初學新語言時（尤其是歐美語系）所遭遇的困境一樣呢！

P.187

What if IBM invested years of research in the project and spent millions on marketing, only to be shown up by a lone hacker working in his basement for a month?

👉 be shown up by 是片語，查到的英英解釋是To outperform or outclass someone; to make someone look unskilled or inadequate by comparison with one's effort or talent

P.188

Salient顯著的，突出的

Watson, however, treated each question as a massive research project. Its process was a lot like searching for the perfect person to hire for a job opening.

👉Watson在對付每一個問題時，就像是進行一場大型研究一樣，或者該說像是在進行一場大型面試，只為了尋找到那一位最完美的候選人（也就是正確答案）。

P.200

Heterogeneous 不均勻的；異質的；各種各樣的；混雜的。

Using a classifier on these candidates would be like trying to fit a square peg into a round hole. It just doesn’t work.

👉 square peg (in a round hole): a person whose character makes them unsuitable for the job or other position they are in; a person or thing that is a misfit; 格格不入

Conductive to =有利於；有利的、有助的、有益的、促成的

P.202-203

And all throughout, …

👉in every part, or during the whole period of time；prep.（表示時間）自始至終；在…期間；遍及…地域；遍及…場所；adv.處處；始終；在所有方面；

DeepQA as if they were the same thing; but they were technically somewhat different. DeepQA is a data-processing engine, and Watson—at least Watson the Jeopardy-playing program I’ve talked about in the past two chapters—was built on top of DeepQA.

DeepQA has nothing to do with the deep learning. The “Deep” in DeepQA refers to deep natural language processing or deep question answering, phrases IBM used to contrast it with simpler approaches to natural language processing, like the methods used in its individual scorers.

P.204

Mishap = bad luck, or an unlucky event or accident

P.205

Financial incentive =經濟刺激；工作獎金

White papers =白皮書（官方報告書）

P.207

It is not being suggested that we should design the strategy in our own image. Rather it should be matched to the capacities and weakness of the computer. The computer is strong in speed and accuracy and weak in analytical abilities and recognition. Hence, it should make more use of brutal calculation than humans.

—Claude Shannon

Pulleys滑輪；滑車

Gears齒輪

Levers槓桿

Feats = something difficult needing a lot of skill, strength, courage, etc. to achieve it

Emulate =效仿，模仿；和…競爭，努力趕上。

P.208

Crank曲柄，曲軸

P.209

Mild-mannered舉止溫文的

Devoured👉

1. to eat something eagerly and in large amounts so that nothing is left;

2. literary to destroy something completely;

3. to read books or literature quickly and eagerly.

在這一段當中，devoured這個字的意思很明顯是第3個，「手不釋卷」。

P.212

Enumerate列舉，枚舉

Pruning = prune：修剪（樹木等）；精簡某事物，除去某事物多餘的部分；

P.214

Branching factor 分支因子

P.218

…this computer would grind to a halt on searches just two levels deeper…

👉grind to a halt = to stop slowly；慢慢停下來，逐漸陷入停頓；to stop or no longer work well

摘自Gavin職場英文的解釋：

Grind to a halt 是什麼意思呢？ grind 除了是「磨」之外，也有「摩擦得嘎嘎響」的意思。 halt 則是名詞「停止」的意思。 Grind to a halt 是「慢慢停頓下來」的意思，你不妨想像一輛老車開得嘎嘎作響，最後慢慢停下來的畫面，就會很有感。（https://www.facebook.com/gavinchangenglish/posts/955617844541424/）

Quaint =古雅的；奇特而有趣的；古色古香的；少見的，古怪的；離奇有趣的；做得精巧的；

Bust =胸像，半身像

P.219

Evaluation function 評估函數

p.223

Singular extension單步延伸

P.224

Culminated =以…告終；達到…的頂點

P.225

Although I’m reluctant to say the answer is a categorical no, there are a few challenges we’d face if we tried to do this.

👉Categorical =絕對的; 明確的

P.226

Reinforcement learning with a neural network gives us a different way to accomplish the same goal. The role of reinforcement learning when playing games is to orient the agent toward states with future rewards by telling it which actions will move it toward those states. Reinforcement learning essentially turns the problem from a search problem (which might be much harder) into a “hill-climbing” problem, where it can move, step by step, toward more promising states.

👉強化學習的定義；強化學習會將一個「搜尋」的問題，轉變成一個「爬山坡」的問題，一步一步地朝更有希望的狀態前進。

P.228

Tesauro’s pitting of the backgammon neural network against itself became a famous story in the field of artificial intelligence, but the method wasn’t widely known outside of the AI and backgammon communities.

👉片語

to be pitted against sb.  與某人競爭;

pit sb/sth against sb/sth 使相鬥；使競爭；使較量

It would take two decades of new ideas and hardware improvements to bring computer Go agents within reach of the best humans.

👉片語

伸手可及的；close enough to be touched or picked up

P.229

…one the most enthusiastic proponents of developing the system and one of its lead researchers, …

👉 proponent = advocate

But Müller and his fellow researchers, unfazed by a challenge, continued to work on the problem.

👉片語

unfazed by = not surprised or worried

不受干擾的；不擔憂的；不為所動的

Juxtaposition = the fact of putting things that are not similar next to each other

P.230

Aficionados =愛好者, devotee

Peculiar =奇怪的，古怪的，罕見的

Elusive =難以描述（或找到、達到、記起）的；困難的

P.231

And so programmers tried and tried for decades, using the typical bag of AI tricks: they…

👉a/the bag of tricks 各種訣竅、方法；全部法寶。

P.233

This is exacerbated by the fact that the game can change quickly: …

👉exacerbated = 使…惡化; 使…加劇

And although I could explain some of this intuition in words, much of it was simply pattern-matching by my subconscious, hunches I had but couldn’t quite put a finger on: …

👉put your finger on sth翻譯：確切地指出（尤指不對勁的地方）

P.235

Sample games are sometimes called “rollouts”…

👉rollouts = the act of making something, especially a product or service, available for the first time;

首次提供（產品或服務）

首次展示（提議，法律或資訊）

P.238

Nefarious =（尤指活動）邪惡的，不道德的。

P.239

Gut wrenching =

1. Making you feel very upset or worried

2. Emotionally painful

And Lee? He gets up and walks out of the room. For a moment it’s unclear what’s happening, but then he re-enters the game room, newly composed, sits down, and plays his response.

👉Newly composed = 重新／再次（變得）鎮靜、沉著的

P.241

plummeted暴跌，急遽下降

Monte Carlo Tree Search = MCTS

Monte Carlo Tree Search was AlphaGo’s solution to both its slow move-prediction problem and its nefarious wrong-move problem.

P.247

They also pitted AlphaGo against itself for millions of games to generate more data to improve the evaluation-function neural network, …

👉這個片語在p.228也出現過一次

片語

to be pitted against sb.  與某人競爭;

pit sb/sth against sb/sth 使相鬥；使競爭；使較量

P.248

Except for its uncanny ability to recognize patterns in the game of Go and to select moves from these patterns—abilities that no doubt were impressive—AlphaGo didn’t demonstrate most of the behaviors we often associate with human intelligence. It couldn’t interact with a fast-changing world. Except for the statistics it aggregated in the upper levels of its search tree, it had no memory of past events; and except for the simulations it ran of how it and its opponent might move, it had no conception of future events. AlphaGo’s creators, like the creators of most of the automata in this book, designed it to solve a narrow problem. For the same reason an airplane doesn’t have wings that flap, AlphaGo doesn’t have a memory or an ability to react quickly to a real-time environment. AlphaGo was engineered precisely to play Go, so it only demonstrates the capabilities required for that.

P.249

Lingo =行話；the vocabulary or jargon of a particular subject or group of people.

P.250

Engrossed =全神貫注的；專心致志的

P.251

Pinnacle =成功；極點；頂點；頂峰；巔峰

P.257

Epoch = 時期；時代

P.258

How could we endow an agent with a memory?

👉be endowed with sth：to have a particular quality or feature；賦予

我們該如何賦予（學習）代理人記憶呢？／我們該如何讓（學習）代理人有記憶呢？

P.259

Sporadically = sometimes, but not regularly or continuously; 偶發地，零星地；

Curricula = curriculum的名詞複數；課程

Pivot = 1. 樞，樞軸；支樞; 2. 【喻】中樞；中心點；中心人物

P.259-260

Go at the university. And the University of Alberta has several of the world’s leading experts on a variety of topics in artificial intelligence, including Richard Sutton, who has been described as the “Godfather of Reinforcement Learning.” One of Sutton’s contributions to the field was the very algorithm that the Atari-playing agent used to learn from its actions—the algorithm it used for off-policy learning.

To make sense of 了解、理解

Draw conclusion from    得出結論

Endeavor = strive; 努力，盡力，竭力；an effort or attempt to do something

P.261

THE FITS AND STARTS OF AI DEVELOPMENT

👉in/by fits and starts：

If something happens in fits and starts, it often stops and then starts again；

時斷時續，間歇的，一陣陣的

Bursting at the seams

👉 If a place is bursting at the seams, it has a very large number of people or things in it；擠得水洩不通，擠滿

Advent =（事件、發明或人物的）出現，來臨，到來；the fact of an event happening, an invention being made, or a person arriving

Derision =嘲笑；嘲弄

Funk = the state of being unhappy and without hope；驚恐；沮喪；懦夫；（因恐懼）避開（某事物）；害怕；畏縮；發出刺鼻臭味；

P.264

Leaderboard =選手積分榜；排行榜

Fungible =易於（與同類或同價物品）交換（或交易）的。可取代的。

P.265

…and progress was rapid in the ensuing years.

👉ensuing =隨後（的）；接著發生（的）

The only bottleneck to how much data these game-playing programs could train on was the time it took the computers to play through their games.

P.266

First, the automata we create in the future will invariably still follow programs. This is a constraint of the media we’ve used to create these automata and the physical laws of the world we live in.

👉media = medium的名詞複數；媒介；途徑；Through or via the medium of sth. =通過某種方法；

Some philosophers have argued that this suggests that machines will never think. My own belief is that humans are machines as well—we’re analog machines—and if we believe that humans can think, then there’s nothing to preclude us from designing digital computers that will also someday think. Rather, it’s inevitable that our machines will someday think, and that they will develop emotions, opinions, and the desire for self-preservation—which will someday conflict with our own.

👉感覺作者這段的結論想法滿有「科幻小說／科幻電影」的氛圍呢。

P.267

Building machines in our image is a human endeavor, and certain qualities of human nature—curiosity, aesthetics, hubris, and vanity, but mostly curiosity and aesthetics—will compel us to continue.

👉Let us make man in our image, in our likeness. = 讓我們按照我們的形象，按照我們的樣式造「人」。

hubris翻譯：驕傲自大。

compel翻譯：強迫；逼迫；迫使, （有時指並非所願地）激起，引發。