INB Lunch-Seminar

Neural Image Processing in Early Vision

Dr. Matthias Bethge
Leiter der Gruppe "Computational Vision and Neuroscience,
Max-Planck-Institut für biologische Kybernetik,
Tübingen

The visual input to the retina consists of complex light intensity patterns. The interpretation of these patterns constitutes a challenging task: for object recognition it is not clear what information about the image should be extracted and in which format it should be represented. Similarly, it is difficult to assess what information is conveyed by the multitude of neurons in the visual pathway. Right from the first synapse, the information of an individual photoreceptor is signaled to many different cells with different temporal filtering properties, each of which is only a small unit within a complex neural network.

Leaving aside the biophysical complexity, it is commonly assumed that the first stages of visual processing implement a filter bank where the neural responses can be modeled as spatio-temporal linear filters plus point-wise nonlinearities. The prevalent tool for characterising the filter properties of these neurons, the spike-triggered average, only allows to describe the stimulus--response function of one single neuron at a time. In order to assess what information is transmitted within a neural pathway, however, it is necessary to get a correct description of the collective behaviour of neuronal populations. Can we find a concise description for the processing of a whole population of neurons analogue to the receptive field for single neurons?

In the first part of my talk, I will present a generalization of the linear receptive field which is not bound to be triggered on individual spikes but can be meaningfully linked to arbitrary response patterns. More precisely, we seek to identify those stimulus features and the corresponding patterns of neural activity that are most reliably coupled. As an efficient implementation of this strategy, we use an extension of reverse-correlation methods based on canonical correlation analysis. We evaluate our approach using both simulated data and multi-electrode recordings from rabbit retinal ganglion cells.

In the second part of my talk, I will address the question what the computational purpose of processing the retinal image with a bank of filters may be. Previous studies have shown that certain receptive field properties can be derived from the goal of redundancy reduction. In particular, the localized, oriented, and bandpass filter shapes of V1 simple cells have been linked to higher-order decorrelation by means of independent component analysis (ICA). Earlier attempts to quantify the difference in coding efficiency between the orientation selective ICA filters and those derived with second-order decorrelation methods yielded differing results for the coding gain of ICA. In a comprehensive study we included all the previous approaches by measuring the expected log-likelihood, the multi-information, as well as rate-distortion curves for both gray-level and color images. Without exception, we find that the advantage of ICA in comparison with second-order methods is very small. We further corroborate and explain this finding by showing that natural images are better fit by a spherical symmetric distribution than by the ICA model. In conclusion, the idea that the goal of V1 simple cells is to provide a less redundant representation of the retinal image does not seem compatible with the assumption of a linear filter bank.

Zeit: Freitag, den 8. Juni 2007, 12 Uhr c.t.

Ort: Institut für Neuro- und Bioinformatik
        Seminarraum (1. OG, Raum 17),
        Ratzeburger Allee 160 (Geb. 64, 1. OG)