Published in Computational and Systems Neuroscience (CoSyNe), (II-91), Feb 2012.
As visual information passes through the visual hierarchy, neuronal responses are thought to become more selective to complex image features, and more invariant to identity-preserving image transformations. This transition has been observed for neural populations in areas V4 and IT (Rust and DiCarlo, 2010), although the precise nature of representations in these areas remains elusive. We examined this hypothesis in areas V1 and V2, using stimuli generated with a model of naturalistic image structure that has been linked to visual representation in V2 (Freeman and Simoncelli, 2011). We generated novel stimuli with model responses matched to those of an original natural texture image. These images are perceptually similar, despite differences in their local features. We assessed how accurately the identity of such images can be estimated from neural populations by recording sequential single-unit responses in anesthetized macaque and training linear decoders on the combined population. To test for selectivity, we asked whether neural populations discriminated better among images generated from the model than among spatial-frequency-matched images which lack the complex features captured by the model. Identification accuracy in V1 was indistinguishable for the two types of image, whereas V2 accuracy was 20% better for the model-generated images. Thus the V2 population is more selective for the features captured by the model than the V1 population. To test for invariance, we measured discrimination performance when the linear decoder was forced to generalize across multiple distinct images matched according to the model. The performance decrement was 50% larger for V1 than V2, indicating that the representation in V2 is more invariant to those features discarded by the model than the representation in V1. Our results demonstrate increases in both selectivity and invariance from V1 to V2, and explicitly link the features of our model to population responses in area V2.