Purchase this article with an account.
W.Y. Kao, S.A. Beardsley, L.M. Vaina; Perceptual learning of motion-pattern discrimination: Psychophysics and computational modeling. Journal of Vision 2002;2(7):73. doi: 10.1167/2.7.73.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
Purpose. A) To study perceptual learning during a motion-pattern discrimination task and to probe for transfer of improvement from trained to untrained stimuli. B) To quantify a neural architecture and computational rules that mediate learning in a physiologically inspired model of motion-pattern processing. Methods. A) In a graded motion-pattern (GMP) discrimination task, random dot stimuli were presented in which the global motion pattern was perturbed relative to a reference motion pattern. Subjects were separated into 2 groups and trained (8000 trials) to discriminate radial or circular perturbations. B) Learning was explored using an interconnected population of MST-like neural units to model psychophysical performance on the GMP task (Beardsley & Vaina, 2001). We tested several physiologically plausible learning rules to modify these lateral connections and shape the resulting population code. Results. A) For each observer, learning was evaluated using an ANOVA across sessions and a t-test of threshold differences between first and last sessions. Learning occurred across subjects and was retained, but did not transfer to the untrained orthogonal motion patterns. B) When we used Hebbian/anti-Hebbian learning rules, the model developed inhibiting interconnections between non-preferred units, resulting in human-like improvements in performance across training sessions. Conclusions. A) The lack of transfer in the GMP task is consistent with previously reported physiological evidence suggesting the existence of specialized motion-pattern detectors in humans. B) The model links these detectors within a functional context to MST, demonstrating how a highly interconnected neural architecture can encode the perceptual information necessary to perform the GMP task. Using simple learning rules, the strength of these connections changed with training, inducing lateral inhibition between non-preferred MST units. This leads to an efficient encoding specific for the trained motions.
This PDF is available to Subscribers Only