Pigeons and perceptrons

摘要

The laboratory pigeon is able to classify images into natural categories such as people, trees, fish, etc. A series of experiments suggested that the pigeon does so by cueing on specific features. There was no tolerance of noise or of perspective distortion, and no recovery of the three-dimensional target. It appeared that the pigeon's capacity could be described by a diameter-limited perceptron.A second series of experiments revealed that feature detection was, in fact, translation invariant: responding extended to any permutation of the feature set. Thus, the pigeon can be more accurately described as extracting the first order invariants from an image class. This first order limitation may be tied to the near-field grain acquisition system, and may not apply to the far-field flight control system.