Invariants

In the AI Story

The discovery that invariants carry information about environmental structure was the mathematical breakthrough that allowed Gibson's ecological approach to become a rigorous alternative to inferential theories of perception. Where the traditional view required the brain to construct three-dimensional structure from two-dimensional retinal data through unconscious inference, Gibson's framework showed that higher-order structural properties — invariants — were present in the ambient array itself and could be directly detected.

Gibson identified several specific classes of invariants: structural invariants of object shape that persist across viewpoints, transformational invariants that specify motion and events, and invariants of the observer's own locomotion that specify egomotion and the layout of surfaces relative to the observer's action capacities. Each class carries specific information relevant to specific classes of action.

The detection of invariants is not an inference from samples; it is the organism's progressive sensitivity to what remains stable across the samples. Perceptual learning, in this framework, is the education of the perceptual system to detect invariants it previously could not pick up. The expert radiologist sees invariants in the texture patterns of X-rays that the novice does not notice. The experienced builder sees invariants in the structure of a codebase — patterns of architectural fragility, signatures of technical debt — that a newcomer cannot detect.

The relevance to AI is subtle but important. Current machine learning systems extract statistical regularities from training data, which in some cases corresponds to invariants in Gibson's sense and in other cases does not. The difference matters because invariants are properties of the structured environment, detectable through active engagement, while statistical regularities are properties of sampled distributions, computable without engagement. Systems that confuse the two may achieve high performance on familiar distributions while failing on novel situations that require the detection of genuine environmental invariants.

Origin

Gibson developed the invariants framework progressively from The Perception of the Visual World (1950) through The Senses Considered as Perceptual Systems (1966), with the mature mathematical treatment in The Ecological Approach (1979). Collaborators including Gunnar Johansson and William Warren developed the framework mathematically and experimentally.

Key Ideas

Persistence under transformation. Invariants are properties that remain unchanged as the ambient array transforms through the observer's motion.

Specification, not inference. Invariants specify environmental structure directly; the organism detects them rather than inferring them.

Multiple classes. Different invariants specify object shape, motion events, and the observer's own locomotion — each relevant to different classes of action.

Detection requires engagement. The perceptual system learns to detect invariants through active exploration; the invariants are in the array, but the detection requires the organism's movement.

The AI test case. Systems that extract statistical regularities from training data may or may not detect genuine invariants; the distinction determines performance in genuinely novel situations.

Debates & Critiques

The mathematical status of invariants remains debated. Strict ecological psychologists treat them as specifiable higher-order variables in the ambient array. Critics argue that what Gibson called invariants are often computational constructs whose extraction requires the kind of internal processing Gibson's framework rejected. The dispute bears on whether contemporary computer vision systems — which extract something invariant-like through deep learning — represent vindication of Gibson or merely a sophisticated reintroduction of the inferential machinery he argued against.