Modeling Orientation Discrimination at Multiple Reference Orientations with a Neural Network

We trained a multilayer perceptron with backpropagation to perform stimulus orientation discrimination at multiple references using biologically plausible values as input and output. Hidden units are necessary for good performance only when the network must operate at multiple reference orientations. The orientation tuning curves of the hidden units change with reference.

Our results suggest that at least for simple parameter discriminations such as orientation discrimination, one of the main functions of further processing in the visual system beyond striate cortex is to combine signals representing stimulus and reference.