Abstract
Unsupervised learning models applied to natural scene statistics almost ubiquitously generate V1-like
receptive fields. To assess further how well these models describe the functional goals of its neural
representation, we applied such a model to the V1 responses of animals raised in environments with unnatural
statistics, namely the recent sustained goggle-rearing experiments of Tanaka [2]. We show that an
over-complete product-of-experts model (POE) [1] captures well many characteristics of V1 simple cells
observed under both normal and goggle-reared input.
Tanaka [2] used optical imaging and electrophysiological methods to show that severely restrictive
striped goggle rearing for many months post-eye opening has extreme effects on neural development. The
percentage of neurons in goggle-reared kittens preferring the orientation permitted by the goggles was
over five times that of neurons in normal kittens along with other more subtle changes.
We simulated the effects of goggle-rearing by training POE, an over-complete extension of independent
components analysis, with inputs consisting of unadulterated natural scenes and/or natural scenes that had
been filtered with (software-defined) goggles. Different proportions of the two inputs were used in order
to model innate mechanisms favoring the statistics of natural scene-like input or incomplete striped
rearing. We applied POE models with differing degrees of over-completeness, the other variable that is
known to exert significant influence over the nature of receptive fields in functional models.
Our results show that there is a significant regime in which model
filters at the goggle-filtered orientation (GO) are over-represented,
together with a relatively even distribution of filters for other
orientations, as is seen in the experimental data. The degree of
orientations selectivity can be quantified by an over-representation
index, ORI=(n oriented at GO)/(n oriented elsewhere) where n is
the number of neurons. In experiment, ORI values ranged from
3.74-12.7 [2] and the results of our modeling spanned a similar
range, with ORI values that scaled with the percentage of GO
stimuli present in the input. Also like goggle-reared neurons, our
model neurons exhibit; a lower proportion of oriented-localized
neural RFs relative to normally-reared kittens; narrower tuning
widths for RFs at GO (characterized by the full width half max
(FWHM) of the tuning curve; and larger and more elongated shape of RFs at GO ([2]; and Tanaka personal communication).
