Alistair J. Bray, Harry G. Barrow
This document describes an activity-based model of information processing in the early mammalian visual pathway. The work has been published previously in an abbreviated form so this report is intended to present a more complete story. The model describes the retina, lateral geniculate nucleus (LGN) and development of simple cells in visual cortex (layer IVc of V1). In the retina we employ a non-adaptive model of on-centre and off-centre retinal ganglion cells (the tonic cells only) that is a non-linear approximation to difference-of-Gaussian processing. The output of this provides input to the LGN where the On and Off channels are kept separate. Here we simulate the effects of local inhibitory lateral interactions; this stage is also non-adaptive. Simple cells in the cortical model receive feedforward excitation from both the On and Off channels projecting out of the LGN. We propose a dual population model in which one population excites close neighbours while the other inhibits all neighbours within a greater area. We simulate this dynamic feedback using an iterative method, and when the network activity is stable we adapt all feedforward weights connecting simple cells to the LGN using a Hebbian (correlation-based) learning rule. We find that when presenting the network with many samples from natural images, the simple cells' feedforward weights adapt to become 'edge' and 'bar' detectors with receptive fields extremely similar to Gabor functions. These edge and bar detectors are orientation slective, and the orientation preference of different simple cells varies smoothly across the cortical surface. When plotting preferred orientation we get 'orientation maps' qualitatively similar to those documented in neurophysiologica literature (in terms of smoothness and singularities). We examine these maps (and others) in terms of their auto-correlation matrix and orientation distribution. Finally we describe preliminary experiments in which the lateral connections within the cortex are adaptive; we find that regions of simple cells develop within which the cells have similar receptive fields but between which the cells have different receptive fields.
This paper is not available online