Lecture 21: Deep Learning

Lecture 21: Deep Learning¶

AIMA Chapter 21 — 1 hour¶

Learning Objectives¶

• Understand feedforward networks

• Apply backpropagation and gradient descent

• Build convolutional and recurrent networks

• Use regularization: dropout, weight decay

• Apply transfer learning

Neural Networks¶

• Perceptron: Linear + threshold

• MLP: Multiple layers, nonlinear activation

• Universal approximation: Can approximate any function

Backpropagation¶

• Forward pass: Compute activations

• Backward pass: Chain rule for gradients

• Update: Gradient descent on weights

Activation Functions¶

• ReLU: max(0, x)

• Sigmoid, tanh: Saturated

• Softmax: Output layer for classification

Convolutional Networks¶

• Convolution: Local receptive fields

• Pooling: Downsampling

• Architecture: Conv → Pool → ... → FC

CNNs: Key Ideas¶

• Parameter sharing: Same filter everywhere

• Translation invariance

• Hierarchy: Low → high level features

Recurrent Networks¶

• Sequence: Process one step at a time

• Hidden state: Carries information

• LSTM: Long short-term memory

Regularization¶

• Weight decay: L2 penalty

• Dropout: Randomly zero activations

• Batch norm: Normalize activations

Unsupervised Learning¶

• Autoencoders: Reconstruct input

• GANs: Generator vs. discriminator

• VAE: Variational autoencoders

Transfer Learning¶

• Pretrain: On large dataset

• Fine-tune: On target task

• Feature extraction: Freeze early layers

Summary¶

• Backprop: Gradient computation

• CNN: Convolution, pooling

• RNN/LSTM: Sequences

• Regularization: Dropout, etc.

References¶

• AIMA Ch. 21

• Russell & Norvig, AIMA 4e, Ch. 21

• Chapter PDF: chapters/chapter-21.pdf

• aima-python: neural_nets.ipynb, deep_learning4e.py

Questions?¶

Next lecture: Reinforcement Learning (Chapter 22)