Using a connectionist mechanism, the dependence of perception on invarient information was investigated. The ability of human perceivers to make accurate judgements about objects which are coming toward them, and comparison to the ability of an associative network to make analogous judgements lead to conclusions both about the model's relation to the mechanism of human perception and the role of invariance in the perceptual process. The human subjects were tested on their ability to make a catching response appropriate to the trajectory of a simulated looming object. This ability was found to be significantly better when the simulation included the optical expansion of the object, which is the information necessary to make the judgement. A linear associator was then trained to make either simple left-right, relative judgements about simulated looming objects, or judge their position at an arbitrary plane absolutely. When trained with stimulus sets which always contained information about the position of the looming objects through their expansion patterns, the model replicated the human data. When the nature of the teaching stimuli was altered to destroy the structure of the information available to the system, it failed to learn at all.
ASJC Scopus subject areas
- Artificial Intelligence