Neural-Network Foundations

Imagine a spam filter built from a single artificial neuron. It gets two clues about an email — say, how many exclamation marks it contains and whether you know the sender — weighs each clue, adds them up, and fires: spam or not spam. Three numbers total: two weights and a bias. Then it gets one wrong. What should happen to those three numbers next? Answer that — really answer it, mechanically, no hand-waving — and you have understood how every neural network on Earth learns.

Four core sessions assemble a working neural network from nothing, starting with exactly that neuron. The perceptron module traces the weighted sum, bias and threshold number by number, then shows the founding idea of the whole field: learning is nothing more than adjusting the weights after each wrong prediction. From there you meet the activation functions whose bend gives networks their power — a single neuron can only ever draw a straight line through its clues, and the bend is what lets stacked neurons draw anything else. You then wire two layers into a feedforward pass you can trace by hand, and finish by scoring the whole thing with a loss function: a single number that says how wrong the network is, and therefore what "getting better" means.

Nothing here is a black box. Every module carries an interactive trace where you nudge a weight or swap an activation and watch the output move — the spam-filter question above is one you will settle with your own mouse, not by taking anyone's word for it.

Why so small a cast? Because everything that follows — the training loop in Deep-Learning Architectures, and eventually the Transformer itself — is these same four parts repeated at scale. The core path asks for no calculus; an optional fifth module (IT/ML lane) reopens the loss functions with the full mathematical treatment for readers who want the derivations.