A Computational Model of Perceptual Learning through Incremental Channel Re-weighting Predicts Switch Costs in Non-stationary Contexts

Abstract:

Associative re-weighting of early sensory representations accounts for temporal dynamics and switch-costs of perceptual learning in a non-stationary environment. 13 human observers discriminated the orientation of peripheral Gabor targets in two filtered noise contexts for 8 one-hour sessions (9600 trials with feedback). The training schedule alternated 2-day blocks of each context. Both discriminability d' and speed improved within and across blocks. A transient switch cost occurred whenever the predominant background orientation changed. Also, for context-congruent targets, accuracy paradoxically decreased slightly with increasing Gabor contrast; for context-incongruent targets, accuracy increased substantially with contrast.

These data impose strong constraints on the possible neurological mechanisms for perceptual learning. In particular, they seem problematic for the widespread hypothesis that it is due mainly to reorganization and sharpening of the early visual representations. Although well documented in other modalities and/or after lesions, this hypothesis has not received much support in intact adult visual areas. An alternative hypothesis attributes the improvement to selective strengthening of the read-out connections to higher cognitive areas.

A detailed computational model demonstrates the coherence of this selective re-weighting hypothesis and its ability to account quantitatively for the empirical results. The model instantiates a number of organizational principles derived from single-cell recordings in the early visual system. It learns via Hebbian potentiation of task-correlated units and suppression of uncorrelated ones.

Slides (pdf)