|
Matthias Bethge
Max-Planck Institute, Germany
Tuesday 9 May 2006
15:00
Seminar Room B10 (Basement)
Alexandra House, 17 Queen Square, London, WC1N 3AR
Linking receptive fields to perceptual inference in the
early visual system
The ability to infer behaviorally relevant information about the environment from the retinal image constitutes an intriguing phenomenon in biological vision whose explanation has proven difficult. The topic of this talk will be to asses in which way sensory neurons in the early visual systems contribute to these inference processes. A long-standing idea is that the sensory neurons are adapted to the statistics of natural images such that the redundancy among the neural responses are minimized. If the computations underlying these responses are described by a linear-nonlinear cascade model, one can ask more specifically which properties of the linear receptive fields are most relevant for achieving this goal.
It has been shown previously that second-order decorrelation alone is not sufficient to explain the receptive field properties of V1 simple cells. Localized, oriented, and bandpass receptive fields, however, are obtained if one additionally minimizes higher-order correlations using independent component analysis (ICA). Despite this success of the redundancy reduction hypothesis, it is not fully understood what the advantage of this additional higher-order decorrelation is. In particular, i will present quantitative results, showing that the additional minimization of higher-order correlations has only a small effect on the total amount of redundancy reduction (less than 5%). In addition, i will present psychophysical data, which indicate that the ICA basis functions of natural images exhibit significant perceptual redundancies. Based on these findings, i will discuss how future work may demonstrate more clearly what the early visual system contributes to visual processing.