Investigating Nonlinearities in the Standard Model of Object Recognition

The Problem: The standard model of object recognition in cortex is a computational model of ventral pathway thought tomediate visual object recognition [5]. It is composed of hierarchical feedforward layers of neuron-like units, performing either (1) template matching with a Gaussian function to increase feature complexity or (2) nonlinear pooling with a maximum operation to increase translation and scale invariance (see Fig. 1a). Using these two nonlinear operations, the model can successfully reproduce the selectivity and invariance properties of the object-selective IT neurons. However, as naturally expected, the neural computation in cortex is variable and not exact. For example, recent electrophysiological experiments have found that complex cells in cat primary visual cortex exhibit behavior ranging from linear summation to maximum [1], and it is also quite arguable that the tuned response profile to various stimuli follows a Gaussian tuning function. Therefore, in order to be a viable model of the biological system, the standard model has to be robust under variation of these nonlinearities. In other words, would the model achieve the similar degree of selectivity and invariance, necessary for object recognition, even when the Gaussian template matching and the maximum operations are made less precise, or less nonlinear?