The artificial neuron, activation functions and the vanishing-gradient problem, forward and backward propagation, gradient descent, loss functions, and why CNNs and RNNs each hit real architectural limits before transformers arrived.
Deep learning is the specific mechanism that makes modern AI work: layers of simple artificial neurons, each doing nothing more than a weighted sum and a nonlinear activation, stacked deep enough to approximate remarkably complex functions. This module builds that picture from the ground up — why ReLU displaced sigmoid in hidden layers, how forward propagation and backpropagation divide the work of prediction and learning, and how gradient descent uses the gradients backpropagation computes to actually update a network's weights against a defined loss function.
It also covers the two architectural families that dominated deep learning before transformers existed — CNNs for spatial data and RNNs/LSTMs for sequences — and is deliberately built to end on RNNs' long-range dependency problem, because that exact limitation is the direct motivation for the self-attention mechanism Module 4 covers next.