Neural Networks & Deep Learning
This course takes the one idea you met at the end of Machine Learning Foundations — a model is a weighted sum trained by gradient descent — and grows it into a deep neural network. You will start from a single neuron, stack neurons into layers, watch a network make a forward pass, then train it live: gradient descent, backpropagation, the effect of depth, and the discipline that stops a network from memorizing instead of learning.
Every module runs entirely in your browser and is genuinely hands-on: you will drag a neuron’s weights and watch its output flip, bend a decision boundary with an activation function, turn a learning-rate dial and see the loss curve dive or explode, step through backprop one gradient at a time, and train a real network on a spiral until it separates the colours. Each module also shows the matching Keras / PyTorch code — read-only, so you can recognize it later, with nothing to install now. Each ends with a short mastery check; pass it to mark the module complete.
The building block
Module 1From a Neuron to a Network
The artificial neuron: a weighted sum plus a bias plus an activation. Activity: drag the weights and bias and watch one neuron act as an AND / OR gate. AI anchor: every deep model is millions of these.
Module 2Activation Functions
Sigmoid, tanh, ReLU — and why a network without them is just a line in disguise. Activity: plot each activation and stack two layers to see non-linearity unlock a curved boundary. AI anchor: ReLU and the deep-learning boom.
Putting neurons together
Module 3The Forward Pass
Stack neurons into layers and push data through. Inputs → hidden layer → output, and the decision regions a small network can draw. Activity: feed points through a live 2-layer network and watch every value light up. AI anchor: inference.
Module 4Gradient Descent in Practice
How a network learns: a loss to minimize, a learning rate to tune, and a training loop. Activity: train a real net live, watch the loss curve fall, and crank the learning rate until it overshoots. AI anchor: every model is trained this way.
Module 5Backpropagation Intuition
The chain rule, working backward to give every weight its share of the blame. Activity: step one training iteration — forward to the loss, backward through the gradients — and watch which weights move most. AI anchor: the algorithm behind all of it.
What makes it "deep"
Module 6What Depth Buys You
Why stack layers at all? Features built on features. Activity: train a net on a tangled spiral and add hidden layers with a depth slider — watch a boundary no straight line could ever draw appear. AI anchor: representation learning.
Module 7Training Real Networks
Big networks memorize. Activity: watch training loss dive while validation loss turns back up, then switch on weight decay, dropout, and early stopping and watch the gap close. AI anchor: every production model fights this battle.
Capstone
Module 8 · CapstoneTrain a Network End-to-End
Put it all together: pick a dataset, set the architecture and learning rate, hit Train, and drive a network from random noise to a confident classifier — reading the loss curve and accuracy as it learns. A synthesis check ties every module together.