Stacking more layers on a limit-order-book or document encoder does nothing useful if activations saturate and backpropagated deltas shrink to numerical dust. This note ties two architectural stabilisers together: normalisation layers that keep per-layer signal scale under control, and residual (skip) connections that give gradients an identity path through depth — so depth becomes a usable degree of freedom rather than a vanishing-gradient trap.
In this lesson
- Learning objectives
- Why depth needs an architectural contract
- Normalisation layers
- Residual connections
- Combining the pieces
Before you start
- Regularization: Weight Decay, Dropout, Early Stopping, and Validation Discipline
- Training Deep Networks: Initialization, Batch Normalization, Dropout, and Adam
The full lesson, with worked examples and exercise solutions, is for members.
Read the full lessonNew to SelfMadeQuant.com? See membership options or explore the curriculum.