Volume 4

Volume 4 - Deep Learning Engineering

Volume 4 builds deep learning from one layer, a matrix multiplication followed by a nonlinearity, up to the transformer blocks inside GPT-2, BERT and Llama 3.1. Sections 1 and 2 train numpy networks on scikit-learn's 1,797 real handwritten digits: a network that learns them, a CNN that reads them as images, and recurrent, LSTM and GRU cells that must remember a digit across blank steps. Section 3 builds attention, the transformer block, BERT and the GPT lineage from published configurations. Every chapter sizes its architecture in parameters, memory and compute, and the closing review uses those numbers to choose between an LSTM, a pretrained encoder and Llama 3.1 8B for one task.