4 lessons
The perceptron, a single artificial neuron
The tiny building block behind every neural network, built from nothing but the dot product and sigmoid you already know.
Backpropagation, training a network with many layers
How the chain rule lets a single algorithm assign blame to every weight in a deep network, all the way back to the first layer.
Training neural networks in practice
What actually changes when you go from plain gradient descent to training a real network - better optimizers, batching, dropout, and batch normalization.
The deep learning landscape, and where to go from here
How CNNs, RNNs, and transformers are all the same perceptron-and-backpropagation foundation, wearing different architectural clothes.