foundation Estimated learning time: 5 h

6.1 From perceptron to multilayer network

You can explain what one neuron computes and why depth helps.

Before:05. Classical Machine LearningUnlocks:07. Natural Language Processing11. The ECE Bridge — Signals, Edge & Embedded AI

One neuron is a weighted sum through a nonlinearity; depth stacks these into hierarchies of features. Universal approximation says width alone suffices in theory, and practice ignores the theorem because depth is what generalises. This opens the deep learning module at the only honest place — the forward pass, done by hand on a tiny network, because without that concrete arithmetic every later abstraction floats free.

Work through these

  • Linear unit, activation, layer

    A single unit computes a weighted sum and then passes it through a nonlinear function, and a layer is many of those side by side. Everything in this module is built from that one arrangement repeated.

  • Universal approximation — and its uselessness in practice

    There is a theorem saying a single wide layer can approximate almost any function, and it says nothing about how to find that layer or how wide it must be. Knowing both halves stops the theorem being misquoted.

  • Depth vs. width

    Stacking layers and widening layers both add capacity, and they do so differently, with depth generally proving more efficient. Understanding why is more useful than the theorem above.

  • Forward pass by hand on a tiny network

    Working a forward pass through a tiny network with pen and paper. It takes ten minutes and it removes the sense that a network is doing something mysterious.

Sign in to keep your progress.

Free resources

Links last checked 29 Aug 2026.

Stuck here?

Ask a mentor. A real person answers, and they can see exactly which topic you're on. Usually within a couple of working days.

Checking your session…

Topics shown in module order.