What Makes a Good Feature?

Perceptual information processing systems, both biological and non- biological, often consist of very elaborate algorithms designed to extract certain features or events from the input sensory array. Such features in vision range from simple on-off units to hand or face detectors, and are now almost countless, so many having already been discovered or in use with no obvious limit in sight. Here we attempt to place some bounds upon just what features are worth computing. Previously, others have proposed that useful features reflect non-accidental or suspicious configurations that are especially informative yet typical of the world such as two parallel lines. Using a Bayesian framework, we show how these intuitions can be made more precise, and in the process show that useful feature based inferences are highly dependent upon the context in which a feature is observed. For example, an inference supported by a feature at an early stage of processing when the context is relatively open may be nonsense in a more specific context provided by subsequent higher-level processing. Therefore, specification for a good feature requires a specification of the model class that sets the current context. We propose a general form for the structure of a model class, and use this structure as a basis for enumerating and evaluating appropriate good features. Our conclusion is that ones cognitive capacities and goals are as important a part of good features as are the regularities of the world.