Inducing discrimination in biologically inspired models of visual scene recognition

To enhance the understanding of human perception and mimic it into an artificial system, several types of graphical models have been proposed that emulate the functionality of neurons in biological neural networks. In this work, we investigate the discriminatory power of two such probabilistic models of vision: a multivariate Gaussian model [1] and a restricted Boltzmann machine [2], both widely used to solve classification problems in computer vision. We quantify the generative ability of these models on standard benchmark data sets and show that neither approach on their own is powerful enough to carry out vision tasks because of the very low discrimination they achieve. There is clearly a need for inducing discrimination by a mechanism that exploits these generative models. We show that the Fisher kernels [3] derived from both the Gaussian and restricted Boltzmann machine can significantly improve the classification performance on benchmark tasks while maintaining the biological plausibility of its implementation [4].