Library · book

Deep Learning from Scratch

Seth Weidman
2019·O'Reilly Media

Source: https://www.oreilly.com/library/view/deep-learning-from/9781492041405/

Build-to-understand, one layer below the language model.

Weidman constructs neural networks from first principles in plain Python — the maths, the computation and the intuition developed together — from multilayer nets to CNNs and RNNs, then shows how it all maps onto PyTorch.

Where Raschka builds a language model, Weidman builds the machinery beneath it: what a layer, a gradient and backpropagation actually are.

For anyone who learns by taking things apart, it turns the primitives every model rests on from black boxes into things you have made yourself.

Central argument

A from-first-principles construction of neural networks in Python: it develops the mathematics, the computation and the mental model together, building multilayer networks, convolutional and recurrent networks by hand before mapping them onto PyTorch. The thesis is understanding by construction — nothing is a black box if you built it.

Critique

Its scope is neural-network foundations up to about 2019: no transformers, no LLMs. It is deliberately narrow and hand-built, which is its value and its limit — pair it with Raschka for the language-model layer and Goodfellow for the deeper theory.

Why it matters for product

For a CTO/CPO who learns by taking things apart, this is the 'build to understand' rung one level below the LLM: it demystifies the primitives — layers, gradients, backprop — that every model, frontier or local, is ultimately made of. Own these and the higher-level books stop being magic.