Abstract
Residual eye movements introduce positional variation of stimuli on the retina in different trials
of psychophysical experiments, a fact usually overlooked in models of perceptual learning. Using
a bisection task as an example, we show that small positional variation changes the structure of
the decision rule from linear to quadratic dependence on neural activities that code the input
stimuli, invalidating linear feedforward models. We propose that a recurrent intra-cortical net-
work, presumably in V1, that pre-processes the stimuli and is tuned through perceptual learning,
is responsible for improved performance. Computer simulations in a network designed for a
specific distance (2-D) between the outer bars in the bisection stimuli confirm this. However,
such networks improve and impair performances for D0 6 D, implying positive and negative
transfers of learning. Psychophysical tests, with bisection stimuli on an analog monitor controlled
by a Macintosh, however, found only positive transfers, ie performance improvements, for
D0 D D=2, based on training at D. No transfer was found from line to dot bisection stimuli,
so learning was indeed perceptual, and transfer cannot be attributed to a general improvement.
