Comparable perceptual learning with and without feedback in non-stationary context: Data and model

A small but persistent response asymmetry favors the background orientation.The criterion control mechanism compensates for most of this natural imbalance, but not all. For context-congruent targets, accuracy (Pcorrect) decreases slightly with increasing Gabor contrast; for incongruent targets it increases greatly. As the congruent "channel" is compromised by noise, the system assigns greater weights to the incongruent one. Coupled with the compressive nonlinearities in the representations, this "off-channel looking" explains the counterintuitive reversal for congruent stimuli. A fully implemented, biologically plausible model takes images as inputs and produces discrimination responses as outputs. The stimuli are first processed by standard orientation and frequency tuned units, divisively normalized. Learning occurs only in the "read-out" connections to the decision unit; the stimulus representations never change. An incremental Hebbian rule tracks the external feedback when available, or else reinforces the model's own response. An a priori bias to equalize the response frequencies stabilizes the model across switches. The excellent quantitative fits to a challenging data set demonstrate that Hebbian channel reweighting, with no change in the early visual representations, is sufficiently powerful to account for perceptual learning. The model handles both the feedback and the no-feedback data with essentially the same parameters, demonstrating that self-supervised learning can successfully compensate for the lack of environmental feedback in perceptual tasks. The main empirical results, and their interpretation, are: Fixed Early Representations + Hebbian Learning