Abstract
We modelled the shape selectivity of single neurons of the midSTS body patch of macaques (Popivanov et al., J. Neurosci., 2014). We employed "adaptive stimulus sampling" (Brincat and Connor, Nat. Neurosci., 2004), producing shapes that were adapted on-line to the response of each single neuron during the recordings. This procedure resulted in a large number of shapes, eliciting a wide range of responses for each neuron. We fit quantitative models of shape selectivity to the responses of each neuron separately (n = 77). We employed cross-validation fitting procedures and the model's performance was evaluated with data that were independent from the training data. By ascending order of performance (median explained explainable variance), we fit 1) Curvature and Angular Position tuning models, adapted from Brincat and Connor (2004), in which the model neuron summates the output of a few subunits, each tuned for a combination of curvature, orientation and position (x-y coordinates) of contour elements, 2) Pixel gray level models that consist of a linear combination of pixel gray levels, 3) HMAX models, which are an implementation of the shallow convolutional network HMAX by Mutch and Lowe (2006) and 4) deep Convolutional Neural Network models (CNNs), specifically Alexnet and VGG19. The deeper pooling and convolutional layers (before the fully connected layers) of the CNNs performed best , resulting in a median explained explainable variance of the responses of up to 77%. These successful models consisted of a linear weighting (estimated by PLS regression) of the units of a single deep layer. Furthermore, the performance of the deeper layer CNN model neurons generalized to reduced stimuli of image parts and were largely invariant to scale, as real midSTS neurons. These data show that deep learning networks are unprecedented successful models of the shape selectivity of single neurons of the macaque midSTS body patch.
Meeting abstract presented at VSS 2017