Unit 09: Dense Neural Networks
The bridge between classical ML and deep learning. Before tackling convolutional or recurrent networks, this unit builds intuition for how neural networks learn at all — forward propagation, loss, and the chain rule applied layer by layer. This unit is lighter on scheduled class time than others; some of the material is woven into the CNNs unit that follows.
Concepts You’ll Learn About
- Perceptrons — the simplest neural unit; weighted sum + activation; the connection to logistic regression
- Multi-layer perceptrons — hidden layers; why depth matters; universal approximation (informally)
- Forward propagation — computing predictions from weights
- Loss functions — cross-entropy for classification; MSE for regression
- Backpropagation — the chain rule applied recursively; computing gradients layer by layer
- Gradient descent — updating weights in the direction that reduces loss; learning rate; local minima
- Activation functions — ReLU, sigmoid, tanh; why non-linearity is essential; vanishing gradients
- Softmax — turning raw scores into a probability distribution over classes
Topics
-
Softmax and output layers — converting network outputs into class probabilities; relationship to logistic regression. View Download Run
-
Activation functions — why we need them, how they affect training dynamics, and how to choose between ReLU, sigmoid, and tanh. View Download Run
Notes
- Backpropagation — worked example and theory — a fully worked 2-2-1 network with hand-traced forward and backward passes.
What’s next
Unit 10 applies these ideas to images — convolutional networks replace the dense layers with learned spatial filters that are far more efficient on image data.