Abstract
To find the neural code for shape in extrastriate visual area V2, we have taken a nonlinear regression approach. We recorded from single neurons in area V2 of awake, fixating macaques while stimulating with a large sample of natural scenes (10,000 to 50,000), flashed rapidly in the receptive field and surrounding area. We then characterized the stimulus-response mapping function for each neuron using a neural network. The neural network determined which natural image features were important for each cell, and revealed nonlinear interactions between these features. It also provided a model that could be used to predict each neuron's response to new stimuli. By applying a visualization procedure to the network, we extracted the stimulus dimensions to which each neuron was tuned.
We find that most V2 neurons show excitatory tuning for a single, dominant Gabor-like feature, as found in previous studies that used sinusoidal gratings as stimuli. However, in almost all cases, this dominant excitatory tuning is modulated significantly by excitatory and/or inhibitory tuning to other orientations and spatial frequencies. There are two major trends. First, spatial frequencies that share the dominant excitatory orientation are usually also excitatory. This suggests that V2 cells pool across spatial frequency to enhance the representation of edges and break camouflage. Second, other orientations are usually inhibitory, and these inhibitory interactions occur across a range of spatial frequencies. This suggests that previously reported tuning of V2 cells for crossed and curved stimuli reflects an interaction between a dominant excitatory peak and complex patterns of tuned cross-orientation inhibition.