15  Neural networks and deep learning

Chapter status: Structural placeholder.

This chapter currently contains the approved heading structure but no reviewed technical content.

15.1 Feed-forward networks

15.1.1 Neurons, layers and activation functions

Content status: Placeholder.

This subsection has not yet been drafted or technically reviewed.

15.1.2 ReLU, sigmoid, tanh, GELU

Content status: Placeholder.

This subsection has not yet been drafted or technically reviewed.

15.1.3 Softmax outputs and cross-entropy

Content status: Placeholder.

This subsection has not yet been drafted or technically reviewed.

15.1.4 Universal approximation and the role of depth

Content status: Placeholder.

This subsection has not yet been drafted or technically reviewed.

15.2 Backpropagation and automatic differentiation

15.2.1 The chain rule on the computational graph

Content status: Placeholder.

This subsection has not yet been drafted or technically reviewed.

15.2.2 Forward and reverse-mode differentiation

Content status: Placeholder.

This subsection has not yet been drafted or technically reviewed.

15.2.3 Practical implementation in modern frameworks

Content status: Placeholder.

This subsection has not yet been drafted or technically reviewed.

15.3 Optimisation of deep models

15.3.1 Stochastic gradient descent and mini-batching

Content status: Placeholder.

This subsection has not yet been drafted or technically reviewed.

15.3.2 Momentum, Nesterov and Adam

Content status: Placeholder.

This subsection has not yet been drafted or technically reviewed.

15.3.3 Learning-rate schedules and warm-up

Content status: Placeholder.

This subsection has not yet been drafted or technically reviewed.

15.3.4 Weight decay, dropout, batch and layer normalisation

Content status: Placeholder.

This subsection has not yet been drafted or technically reviewed.

15.3.5 Data augmentation for images

Content status: Placeholder.

This subsection has not yet been drafted or technically reviewed.

15.3.6 Early stopping and validation monitoring

Content status: Placeholder.

This subsection has not yet been drafted or technically reviewed.

15.4 Convolutional neural networks

15.4.1 The convolutional layer and parameter sharing

Content status: Placeholder.

This subsection has not yet been drafted or technically reviewed.

15.4.2 Pooling, stride and receptive field

Content status: Placeholder.

This subsection has not yet been drafted or technically reviewed.

15.4.3 Padding conventions and output-size arithmetic

Content status: Placeholder.

This subsection has not yet been drafted or technically reviewed.

15.4.4 Feature hierarchies and interpretability

Content status: Placeholder.

This subsection has not yet been drafted or technically reviewed.

15.5 Canonical architectures

15.5.1 LeNet-5, AlexNet, VGG

Content status: Placeholder.

This subsection has not yet been drafted or technically reviewed.

15.5.2 GoogLeNet / Inception

Content status: Placeholder.

This subsection has not yet been drafted or technically reviewed.

15.5.3 ResNet and residual connections

Content status: Placeholder.

This subsection has not yet been drafted or technically reviewed.

15.5.4 DenseNet, MobileNet, EfficientNet

Content status: Placeholder.

This subsection has not yet been drafted or technically reviewed.

15.5.5 Vision transformers and hybrid models

Content status: Placeholder.

This subsection has not yet been drafted or technically reviewed.

15.6 Training strategy for industrial data

15.6.1 Transfer learning and fine-tuning

Content status: Placeholder.

This subsection has not yet been drafted or technically reviewed.

15.6.2 Small-data regimes and few-shot learning

Content status: Placeholder.

This subsection has not yet been drafted or technically reviewed.

15.6.3 Class-imbalance countermeasures

Content status: Placeholder.

This subsection has not yet been drafted or technically reviewed.

15.6.4 Reproducibility: seeds, versioning, provenance

Content status: Placeholder.

This subsection has not yet been drafted or technically reviewed.