Neural Population Codes in Multi-Dimension

Vision is a sense that humans and most animals rely the most onto. Although this sense seems pretty trivial, the visual stimulus is composed of many different dimensions which makes it very complex to decompose, analyze and understand. Scientists state that roughly half of our brain is used for vision processing, it is therefore a very interesting and intriguing area for researchers as it can provide a first basis and good insight towards understanding the brain. As a result, many psychophysical studies have been performed in each of these dimensions, in order to find the physical boundaries to which the humans and animals are constrained Burbeck and Regan [5]. From these measurements, much work has been done to reproduce these behaviors using computational and mathematical models of the neural system. Such models based on our knowledge of neurophysiology, were developed such as to reach the same performance measured in psychophysical experiments. The goal was to provide insights, hypotheses and information about how the brain encode and decode the visual stimulus. These models however, held many strong assumptions such as if the model was dealing with only a single dimension of the visual stimulus. While making the model easier to implement it does yield puzzling or fairly inaccurate and inappropriate results for direct comparison to the performance of the brain. On that account, our project aimed to study existing computational models of the striate cortex in multiple dimensions Zohary [27], and extend them with further experiments and appropriate assumptions. As a result of our implementation, we found that 3000 to 7000 neurons were necessary to account for the measured psychophysical performance if the neurons were selective to two dimensions (i.e. Orientation and Spatial Frequency). We also found that for each dimension added to the selectivity of the neurons, the population size needed to grow by a factor of 1.45 times, in order to keep the same level of accuracy (assuming that the following dimensions are similar to orientation). Further experiments assessed the efficiency of different decoding strategies in single and multiple dimensions, the effects of different parameters in the encoding phase leading to optimal results and finally we tested hypotheses towards an explanation of the observed gains in performance during attention demanding tasks.