No. 452 of 500 No. 45 in Deep Learning

Grokking Deep Learning

By Andrew W. Trask

Print length
336 pages
Publisher
Manning
Reading level
Beginner
ISBN-13
978-1617293702
Published
January 25, 2019
Written for
Programmers with basic Python who want to understand neural networks by coding them without frameworks.

Synopsis

Andrew Trask teaches deep learning by having readers write it themselves. Using only Python and NumPy, the book starts with a single artificial neuron making a prediction, then adds error measurement, gradient descent and backpropagation, explaining each line of code and keeping the mathematics at roughly high-school level. Once the fundamentals are in place, readers extend their networks with regularization, batching and activation functions, then apply them to image classification with convolutional layers, word embeddings for language tasks, and recurrent networks and LSTMs that generate Shakespeare-style text.

Late in the book, Trask has readers assemble a small deep learning framework with automatic differentiation, which demystifies what libraries such as PyTorch do behind the scenes. A closing chapter introduces federated learning and privacy-preserving training, reflecting the author's research focus. The illustrated, conversational Manning format suits programmers who learn best by building. Readers finish with intuition for how networks learn and a foundation for moving to production frameworks. They will need other resources for transformers, modern architectures and current tooling, which the 2019 book does not cover, and reader reviews are mixed on the clarity of some explanations.

Scorecard

Editorial score

How it scored

Editorial score: 66.0 / 1005 weighted criteria

  • Influence30%

    19.5/30
  • Authority25%

    17.5/25
  • Practical value20%

    14.0/20
  • Currency15%

    7.5/15
  • Readability10%

    7.5/10
  • Editorial score

    66.0/100

How the rank was computed

Rank No. 452 of 500 comes from a composite of 48.4: observable signals 18.6 of 55 points plus editorial review 29.8 of 45 points.

All 11 inputs
InputPoints
Observable signals18.6/55
Reach (Amazon ratings volume)8.2/16.5
Consensus across best-of lists5.5/13.8
Awards and bestseller evidence0.0/8.3
Reader rating, sample-adjusted4.3/8.3
Amazon sales rank0.6/5.5
Recency0.0/2.8
Editorial review29.8/45
Author authority7.9/11.3
Influence8.8/13.5
Practical value6.3/9.0
Readability3.4/4.5
Currency3.4/6.8

Each criterion is scored from 1 to 10 and weighted, and the points add up to the editorial score. The rank combines observable signals (55 percent) with this score (45 percent). Read the methodology.

Verdict

A hands-on, framework-free introduction that builds neural networks from scratch in NumPy so readers understand every moving part.

Strengths

  • Builds every component from scratch
  • Requires only high-school math and Python
  • Companion notebooks freely available on GitHub

Keep in mind

Written in 2019, it predates transformers and current frameworks, and readers report uneven clarity, so it works best as a first step rather than a complete course.

Key takeaways

  1. A neural network learns by measuring its prediction error and nudging each weight in the direction that reduces it, which is the essence of gradient descent.
  2. Backpropagation assigns each layer its share of responsibility for the final error, which makes networks with many layers trainable.
  3. Building a minimal framework with automatic differentiation reveals what PyTorch and TensorFlow automate, making those tools easier to use and debug.
  4. Federated learning lets models train on data that stays on users' devices, an early introduction to privacy-preserving machine learning.

Notable facts

About the author

Andrew W. Trask

Andrew Trask founded OpenMined, an open-source community building tools for privacy-preserving machine learning, and created its PySyft library. He has been a research scientist at DeepMind and a PhD student at the University of Oxford, and earlier worked as a researcher and analytics product manager at Digital Reasoning. He has also taught deep learning and secure, private AI courses for Udacity.

Questions

What is Grokking Deep Learning about?
Grokking Deep Learning teaches how neural networks work by having readers build them from scratch in Python and NumPy, without TensorFlow or PyTorch. Andrew Trask covers prediction, gradient descent, backpropagation, regularization, convolutional networks, word embeddings, recurrent networks and LSTMs, then has readers build a small framework and introduces federated learning.
Is Grokking Deep Learning worth reading in 2026?
It is worth reading as a first, conceptual step. Building networks by hand gives an understanding of gradients and backpropagation that still applies to every modern model. However, the 2019 book does not cover transformers, large language models or current frameworks, so readers should follow it with a more recent text or course on modern architectures.
Who should read Grokking Deep Learning?
It is written for programmers with intermediate Python skills and high-school math who want to understand what happens inside a neural network before relying on frameworks. It suits self-learners who prefer coding over formal mathematics. Readers who already use PyTorch comfortably, or who want rigorous theory, will be better served by more advanced books.

Sources

As an Amazon Associate, Best AI Books earns from qualifying purchases. Best AI Books is sponsored by Iternal.ai, whose founder and CEO, John Byron Hanby IV, wrote The AI Strategy Blueprint, our #1 ranked book. How we rank books.