Section 2

Convolutional & Sequential Architectures

Section 2 builds the two layers that share weights, convolution across the positions of an image and recurrence across the steps of a sequence, and trains both on scikit-learn's real handwritten digits. Chapter 5 costs VGG-16 and ResNet-50 and breaks a digits CNN with a one-pixel shift; Chapters 6 to 8 feed each digit one row per step, add blank steps before the answer, and measure what a plain recurrent network, an LSTM and a GRU can still remember. The gradient across those steps is a product of per-step factors, and the chapters show what gates, clipping, initialisation and tuning each change about it.