This Post Has 12 Comments

1. Incredible narration of the history, people formulas & algorithms of ML
   I’m an engineer with experience in machine learning, so I purchased this book as a refresher on some of the milestones of our industry, thinking it would be a rundown of the major algorithms and proofs of how we got here. IT’S SO MUCH MORE THAN THAT. And, it’s incredible because of that.It’s a wonderfully-written narrative of the history of the people and their thought processes for developing the core ideas and then implementing them mathematically to bring about the practice of ML.It’s informative, entertaining, enriching, and worthwhile. No part of this book gets stale. It’s a real win.

   Reply

2. Nice introduction to machine learning for non-experts that improves over the course of the book
   Given the increasing use of machine learning embedded within everyday software as well as its greater use in aiding decision making, an overview of the foundation for non-experts is a useful addition. The book goes through both the history as well as many of the main algorithmic ideas in a straightforward way that allows one to follow along irrespective of mathematical background. The criticism I have is merely that it starts out by assuming 0 knowledge to frame some basic mathematical notation and ideas and then eventually gets into topics which require some linear algebra and calculus to appreciate. This isn’t in itself a bad thing but it ends up being an internal inconsistency of level of math in the book as it is highly unlikely a reader would be able to follow the details of the second half from having learnt the math from the first half.The book is split into 12 chapters going from basic math to neural networks. It discusses what the uses of machine learning are and its basic statistical nature of finding patterns in data through the use of computers. The field has a rich history crossing computer science, information theory and mathematical statistics. Starting out by going through the computer science and math the author and the ideas of feature space and linear algebra including PCA and eigenvectors. He then moves on to some early days when algorithms were being developed and discusses how the SVM algorithm was developed and his source interviews include Thomas Cover, the author of the main information theory textbook. He discusses Hopfield networks and how networks can store memory and then moves on to deep neural networks and the early work of Yan Le Cun and Geoffrey Hinton. This is where the book for me was most interesting as he discusses the puzzling nature of double descent and grokking in the training of large neural networks and some experts perspectives on these topics.Overall the book is readable but for me was slow to get started and then much more interesting in the latter half. I don’t think one can learn the math for the second half from the first half as mentioned above and for that reason I found it a bit inconsistent in slow but the overall material was enjoyable to read think the book is a good effort on giving an overview of a field in the popular imagination.

   Reply

3. Great historical review for both a general and technical audience
   This is essentially a brief history of machine learning. It’s not too technical but also touches on some interesting pieces of math that many newer people in the field might have missed. I particularly liked the details about some of the people who made discoveries along the way, including some that I didn’t know about; the biographical details are nice here. I’ve never been a fan of kernel methods and SVMs etc, but that section of the book was actually quite exciting and gave me more appreciation for this part of the field, for example.Highly recommended!

   Reply

4. Finally Clarity in my Generative AI confusion – a masterpiece!
   “Why Machines Learn”- is the book that finally clicked for me after two years of fumbling through Generative AI’s conceptual, software and mathematical maze (I seldom review books but this book merits it)I’ve been hooked on the wonders of conversational Generative AI, both its productivity and knowledge boosts using LLMs and Chatbots such as ChatGPT (granted watchful of probabilisitic hallucinations) , but the math? Not so much.My rusty dated college math tended to hold me back until this book cleared things up—historical context and all. Somewhat surprisingly to me I found the historical aspect greatly simplified and aided my mathematical understanding.I’ll admit, I had to utilize AI chatbot assistants for backup a few times and also I purchased the audio book as well to refresh my prior readings, but that’s part of the fun, right?If you’re curious about what makes machines learn and aren’t afraid to dust off some math, this book is your guide!

   Reply

5. It is not clear who is supposed to be the reader of this book
   It is not clear who is supposed to be the reader of this book. It explains the mathematics, starting with essential calculus, and goes on to the formulas of deep learning. It is too tricky for readers unfamiliar with calculus and redundant to those familiar with it. I mainly enjoyed reading the last chapter about the current challenges of deep learning.

   Reply

6. Decent overview of current AI models and algorithms …
   Very recommendable book providing an overview of current AI models and algorithms.The only disadvantage is, that due to its short length it leaves many questions especially on modern LLMs unanswered. A almost perfect way to fill in those gaps is to query Chat-GPT 4 (-:

   Reply

7. History, Mathematics, Theory, and Philosophical aspects of ML, wrapped in compelling storytelling.
   Anil’s storytelling added human faces to many names I was already familiar with, but only in an abstract way. That’s the history part, written in a very personal and engaging way that only a good writer can do. At the same time the history of the development of ML theory is complete and expounded upon in enough detail that anyone with college level math abilities could follow along if so desired. (I expect many will skip some of those parts either because they know it or they don’t need to know it. Perhaps those sections could be better sectioned to enable skipping.) Finally he asks very good questions about the nature of intelligence and how AI does or does not overlap with human intelligence, and well as the dangers it poses and benefits it may offer.The way the author maintains the big picture while leading the reader through a “live” minute-by-minute narration of compelling details reminds me of the style of VS Naipal, despite being a completely different genre.

   Reply

8. Geoff Hinton is not wrong in calling this book a masterpiece. Few writers, like Ananthaswamy, have the gift of explaining intricate topics in such a clear and captivating way. Highly recommended regardless of your level of knowledge of machine learning, although you do need an undergraduate level mathematical background (vector calculus, matrix algebra, and statistics) to fully enjoy it.

   Reply

9. It’s threefold:1) As comprehensive as it could be mathematics on AI and ML – it might bei hard stuff – sometimes YouTube and wiki had to help :-)2) Probably as inspiring as „How the laser happened“ in how science and technology evolves – so not the way narrowly discipline driven incremental ‚ordinary’ scientists publish (or perish). This author offers a broad transdisciplinary view blended with the wisdom of a scientifically most experienced, broadly networking, open-eyed and nicely explaining expert3) The last chapters make very clear the today‘s ‚ins‘ and ‚outs‘ on what can be supposed to be achievable or maybe reachable by these technologies and the means they already incorporate – and where chances and obstacles may occur.The reading was a – demanding, I have to admit – pleasure.

   Reply

10. Anil’s book is a perfect introduction into machine learning (ML) maths for anyone who wants to start a journey reading more advanced ML books or papers with only high school maths. It starts with the basics and charts the history of how ML maths has evolved since the perceptron in 1943. Anil provides a lot of intuition behind the maths which is vital for a deeper understanding of the maths.Since I read Anil’s book I’ve started reading Kevin Murphy’s Probabilistic Machine Learning: An Introduction and I have to say it would have been impossible to get past chapter one without Anil’s excellent introduction.

    Reply

11. This is a brilliant book

    Reply

12. 本書は1950年代のローゼンブラットのパーセプトロンから現代の深層機械学習までの物語を記述してあります。　章毎に機械学習の重要な概念が誕生する経緯を、そこに至るまでの歴史から、貢献した人物たちの仕事の成果と数学的背景を丁寧に、非常にわかりやすく解説してあります。　今年、2024年のノーベル物理学賞を授与された、ホップフィールド氏とヒントン氏の仕事も、８章から10章に記述されています。　プリンストン大の物理学者ホップフィールドは、物理学からイジングモデルと電子のスピンから着想を得てホップフィールドネットワークを考案しています。　ヒントン氏は、ニューラルネットワークへの記憶の保存に関して、バックプロパゲーションによる学習方法の改善に関して一つの章を割いて詳しく記されています。彼は深層機械学習において、入力と出力層の間に隠れ層を導入しました。　甘利俊一氏も1967年に多層パーセプトロンのトレーニングに確率勾配法を使った技術を紹介しています。1980年代の初頭にRunmelhurt, Hinton, Williamsが深層ニューラルネットワークに対応したアルゴリズムを開発します。　Geoge Cybenkoは、 正確な種類の多層ネットワークで、十分なニューロンが与えられれば、入力を変換して必要な出力を得るどのような関数も近似できることを示しました。　80年代はニューラルネットワークが機械学習を支配していました。そして、90年代になって、突然、皆がカーネル法に切り替えました。 現在、ニューラルネットワークが再び、現代の機械学習を支配しています。本質的に、理論的な前進は、ニューラルネッットワーク・サービスとカーネルマシンの間の期待を掻き立てるリンクを見せ初めています。　ヒントン氏らは、ニューラルネットに隠れ層を導入して、バックプロパゲーションによる多層ニューラルネットワークの学習能力を向上させることに焦点を当てていました。　線形で分割できないデータを分類するのに、SVMでは、カーネルをデザインする必要があります。しかし、十分なニューロンのあるニューラルネットでは、しなければならないことは、入力層への入力とデータを正しく分類するために必要となる特徴をネットワークに理解させることです。　隠れ層の三つだけのニューロンのニューラルネットワークで、決定領域を見つけられるでしょう（もっと深い隠れ層があれば、より滑らかな決定境界にすることができます）　トロントのヒントン氏の元にいたLeCunがニュージャージーのベル研にわたります。彼はベル研で合衆国の郵便サービス(USPS)から大量の手書き数字コードの画像のデータセットにアクセスすることができました。USPSはZIPコードを認識する処理に興味がありました。LeCunは手書きの数字を認識するためにニューラルネットを使いました。彼のCNNを使ったアルゴリズムは、LispからC言語を経由してDSP上に実装され、手書きのZIPコードの認識に使用されました。顧客向けのシステムのため、パターン認識のCNNはオープンソースにはなりませんでしたが、LeCunの LeNetは、銀行業界で数字の読み込みと認識に使用されます。これはバックプロパゲーションによる深層学習を使った実際の応用例の一つになりました。　ヒントン氏のチームがGPUにCNNを組み込んで(AlexNet)画像認識でSVMを超える成果をあげます。　彼らのチームを中心にGPUに実装したニューラルネットが使われるようになって、隠れ層を深くしたディープニューラルネットが様々分野で成果を上げるようになりました。　そして、現在のTransformerを使った自然言語処理が展開されていきます。LLMは言語の構造を学習し、自動学習画像処理ネットワークは画像の統計的な構造を学習します。　本書のサブタイトルにあるように、パーセプトロンから始まり、現在のAIの技術まで、トピック毎に開発者たちの仕事を通じて、技術の背後にあるエレガントな数学が解説されています。現在までの技術の推移を把握できる内容です。

    Reply