Overview

A wheel feature alone does not classify a car. Neither does a window feature. A model combines several earlier features, suppresses conflicting evidence, and creates a later car representation. That composition is a circuit.

Circuit

A circuit is the subset of a model's features, weights, and nonlinearities that implements a particular computation from earlier representations to later ones.

Dog features connected by weighted edges into a circuit
Dog features come together into a dog circuit. [Source 1] [Source 2] [Source 3] [Source 4]

Build a Curve-Detection Circuit

Change the early edge detectors and step through the computation. Excitatory weights add compatible evidence; an inhibitory weight subtracts an opposing orientation.

1. Input evidence

left edge

0.85

right edge

0.75

opposing

0.20

curve = 0.65(0.85) + 0.65(0.75) - 0.70(0.20)

curve

0.90

circle client

0.97

A circuit explanation names the features, the signed connections between them, and the computation those connections implement.