Learning path
Next moduleModule 11: Gradient descent and backpropagation
Module 10
Neural-network foundations
Neurons, layers, weights, biases, activations, and tensor shapes, the building blocks of every deep model.
- 1. The neural-network forward passCompute a multilayer perceptron's output by hand, track every tensor shape, and see why nonlinear activations are what make depth useful.40 min
- 2. Loss functionsPick the right loss for regression, classification, imbalanced data, and embeddings, compute each one and its gradient by hand, and explain why cross-entropy beats squared error for classifiers.45 min
- 3. Intro to PyTorchBuild a tiny autograd engine, then use PyTorch tensors, autograd, nn.Module, and the five-line training loop to train a network on CPU, and recognize the bugs that loop invites.50 min
A neural network composes linear transformations with nonlinear activations. This module follows the numbers and shapes through a forward pass, then the loss that scores it, then PyTorch, the framework that turns both into a training loop.
| # | Lesson | You will be able to |
|---|---|---|
| 1 | The forward pass | compute a small MLP's output by hand and check every shape |
| 2 | Loss functions | choose and compute MSE, Huber, cross-entropy, focal, and contrastive losses, and explain why cross-entropy trains classifiers better than MSE |
| 3 | Intro to PyTorch | build a tiny autograd engine, then write a PyTorch training and evaluation loop on CPU |
Interactive lab: Forward pass
Code: code/10-neural-networks/