My Reading List for Deep Learning

This is a curated list of what I would recommend as resources for learning about various aspects of deep learning, heavily inspired by this Github repository, although based on my own personal experience.

This reading list is relatively long, and I don’t proclaim to have read every single word on every single page. However, I am a firm believer of developing a good foundation: given how expansive the current state of deep learning is, if you’re starting from scratch there is a lot you have to catch up with. My advice is to take everything in strides, and learn what you need to when you need to; this is inevitable, but not insurmountable!

Building a good foundation

Whether you like it or not, deep learning requires a significant amount of background knowledge in both linear algebra and statistics; you need a good solid foundation before you can build a mansion. The former is a requirement for understanding the core mechanics behind every model, and developing a good intuition for linear algebra can provide you insight into some of the tricks involved for some models (e.g. inner product decoders, the inception architecture). Basic understanding of the latter is required for some of the simpler tasks, for example classification. As we get to more complicated problems, a background in Bayesian statistics is extremely helpful: these ideas form the backbone for probabilistic modelling, which is used for generative models—models that create new data based on what it has learnt.

For linear algebra, I don’t actually recommend a mathematics textbook. Instead, I recommend Mathematics for Quantum Chemistry; don’t let the title throw you off, because the first few chapters in what is already a very short book gives you a quick primer and a good reference for properties of linear algebra. Quantum chemistry actually uses a lot of the same machinery as deep learning (there’s a lot of matrix multiplication and factorization), all of which was designed to solve approximations to the Schrödinger equation. You will learn about expressing concepts as basis functions, projections, and solving linear equations. The book is published by Dover, and so is extremely cheap to pick up and good to have around!

For statistics, I generally avoid typical university textbooks that focus on hypothesis testing (i.e. \(p\)-values) that you might find common in Psychology and Biology. The premise behind a lot of these ideas are “frequentist”, and in my humble opinion you are much better off thinking like a Bayesian statistician instead (although not too much in fear of being paralyzed by uncertainty). For this reason, I recommend Bayesian Data Analysis by Gelman, Carlin, Stern, and Rubin, and for a more applied book, Statistical Rethinking: A Bayesian Course with Examples in R and Stan by McElreath. The former in particular sets you up to frame any problem in terms of likelihoods, and provides case studies to understand how Bayesian statistics can help us solve real-life problems and understand the role of uncertainty.

Putting up the framework

There are two specific resources I would recommend that will set you up for good: Deep Learning by three giants of the field: Ian Goodfellow, Yoshua Bengio, and Aaron Courville, and Andrew Ng’s course on The former provides an extremely solid basis and theoretical underpinnings of the basics of deep learning, while Andrew Ng’s course is more pragmatic, teaching you how to implement these models from scratch. A good book to accompany Andrew Ng’s course is François Chollet’s Deep Learning with Python. In both cases, there is a significant focus on Tensorflow and Keras (for obvious reasons), although learning from Deep Learning should provide you enough abstraction to implement many of the basics.

In essence, the combination of these three materials is sufficient for you to start playing around with deep learning models. Of course, this is admittedly easier said than done because there’s a significant amount of reading, but your future self will thank you!

At this point, many of the latest concepts of deep learning come from academic papers: unlike many other fields, virtually all of the material is available without a pay-wall. We are fortunate that the machine learning and deep learning researchers tend to upload a significant amount of their papers onto ArXiv, which is not true for other disciplines (e.g. Chemistry, Physics, Astronomy, etc.). I recommend finding something you’re interested in solving, and start working towards reading papers that provide solutions to those problems. For example, if you’re working with images, take a look at convolutional models: AlexNet, LeNet, Inception, to name a few (in that order). If you work on numerical/sequential data, check out recurrent neural networks. This Github repository provides paper highlights up until a few years ago, and covers the more seminal papers for a lot of the current state-of-the-art. If I don’t mention one of those papers, it’s probably going to be in that repository.

In the following sections, I’ll be discussing more specific applications that are not always systematically covered or make it into the mainstream media, but are (I think) incredibly cool. This is my idea of a one-stop shop for more forefront papers, that are not always related to one another.

Bonus material: This ArXiv paper provides a fairly comprehensive historical overview of deep learning, dating back to ideas from the early 20th century.

Learning dynamics and capacity

One thing that I haven’t found many posts or articles about is the general idea of how much capacity neural networks are: it’s not a straightforward question to answer, and the literature is actually quite diverse on this matter. While most people might dismiss as this “too theoretical”, there are important implications to be learned by understanding how neural networks retain what information. I find this area quite interesting, because it certainly adds an “organic” component to an optimization problem.



Autoencoders are a neat class of models that try to learn to extract useful features in an unsupervised manner. The basic gist is an encoder model produces an embedding that can be used by a decoder model to reproduce the inputs, and by doing so, learns to essentially compress the important parts of an input into a small feature vector. This blog post provides a comprehensive overview of variational autoencoders.


While variational autoencoders are cool, they are typically limited by the fact that diagonal Gaussians do not make very good approximate to true posteriors in many (maybe most) cases. Normalizing flows are a way to transform simple, easily computable distributions into more expressive ones by means of a series of invertible transforms.


Probabilistic Models

The Bayesian Data Analysis book should provide a good foundation for this section: despite the section title, the focus is more on capturing model uncertainty, à la Bayesian statistics. Uncertainty quantification is an essential part in rational decision making, adding to the overarching theme of “making AI trustworthy” for policy making, self-driving cars, all that jazz. This section is by no means comprehensive yet, and I intend to expand it more.


Reinforcement Learning

The hot topic for deep learning, having neural networks teach themselves how to solve problems through trial and error. This section is a little sparse for me right now, but I will get to populating it soon.


Graph models

Graph theory is a way of modelling diverse problems: for example, social networks, circuitry, and structured data, and of course neural networks. There is also a rise in popularity of probabilistic graphs, because of how easy they can potentially make understanding causal effects. In recent times, many of the mainstream ideas in deep learning, such as convolutional and generative models, have found analogous derivations in the graph neural network literature.



For now, these are the resources I would go to for my deep learning fix. I’ll try and keep this post updated periodically, however you should also be doing your own finding! One of the easiest ways would be to go through ArXiv, and find papers that you find interesting. Given how expansive it is, and the fact that tens to hundreds of new discoveries are being reported every week, my recommendation again is to dive into specifics as you need to solve different problems. It’s very unlikely that you will be able to keep on top of everything, and for your own sanity and mental well-being you should deal with these papers and new ones at your own pace!

Feel free to reach out to me if you have questions, or if you think I missed something and I should add this to the list! Also, please let me know if this helped you out at all!

© 2020 Kin Long Kelvin Lee. All rights reserved.

Powered by Hydejack v8.5.1