Concept library/Architecture
02
6 minute interactive lab

Why neural networks need hidden layers

Why can’t one neuron solve every pattern?

YOU’LL LEAVE KNOWINGDepth composes simple features into useful representations.
THE EXPERIMENT

Turn a straight line into a useful shape.

Move the controls. Watch the internal state respond. The explanation follows the behavior—not the other way around.

Conceptual simulationThe behavior is deliberately simplified and labeled; it teaches the relationship, not a trained production model.
ARCHITECTUREAdd hidden layers

One linear boundary cannot separate the curved groups.

test accuracy52%
learned decision boundary
THE MENTAL MODEL

Depth composes simple features into useful representations.

A layer does not magically add intelligence. It gives the network another chance to recombine simple signals. Successive layers can build edges into curves, and curves into concepts.

IN DEVELOPER TERMSoutput = layer₃(layer₂(layer₁(input)))
GO DEEPER

Check the model against the source.

Connectionism optimizes for intuition, then points you to the rigorous treatment.

  1. Dive into Deep LearningMultilayer Perceptrons
  2. Deep Learning bookDeep Feedforward Networks
NEXT CONNECTION

Why introduce a bend into the network?

Activation functions

Continue to lab 03