Neural networks from scratch, no black boxes allowed

What it does — This repository holds the Jupyter notebooks for Andrew Trask’s book Grokking Deep Learning. Each file maps to a chapter and walks through concepts from forward propagation and gradient descent to CNNs, LSTMs, and federated learning. The README itself is sparse: just a linked list of notebooks, a FloydHub run button, and a 40% discount code for the Manning book.

The interesting bit — The notebooks progress from single-weight prediction and basic gradient descent to building a toy automatic-differentiation engine, then onward to CNNs, LSTMs, and federated learning. It is the educational equivalent of learning to drive stick before you are trusted with an automatic transmission.

Key highlights —

• Chapter-by-chapter notebooks spanning prediction, backpropagation, regularization, activation functions, CNNs, word embeddings, RNNs, and federated learning
• Chapter 13 walks through building a miniature deep-learning framework via automatic differentiation
• Includes less-common introductory topics: exploding gradients, LSTMs, and federated learning
• README offers no dependency information, environment details, or prose descriptions of the notebooks
• Chapter numbering is erratic: Chapters 1, 2, and 7 are absent, while Chapter 14 hosts three unrelated notebooks

Caveats —

• The README is a bare link list with no setup guidance, dependency list, or notebook descriptions
• Chapter numbering is inconsistent (three different “Chapter 14” notebooks, and Chapters 1, 2, and 7 are missing)
• No indication of which Python or library versions the notebooks expect

Verdict — A solid resource if you are learning deep learning and want to see the wiring under the board before you touch a framework. Avoid if you need a self-contained reference or a curriculum that works without the book.