CNS*2020 Online has ended
Welcome to the Sched instance for CNS*2020 Online! Please read the instruction document on detailed information on CNS*2020.
Back To Schedule
Monday, July 20 • 7:00pm - 8:00pm
P219: Interpretable modeling of neurons in cortical area V4 via compressed convolutional neural networks

Log in to save this to your schedule, view media, leave feedback and see who's attending!

Feedback form is now closed.
Zoom Link: https://ucsf.zoom.us/j/98471222466?pwd=UllvdG5USmYwdkE2VUJXcnBkTGl0dz09

Reza Abbasi-Asl
, Bin Yu

Characterizing the functions of neurons in visual cortex is a central problem in visual sensory processing. Along the ventral visual pathway, functions of the neurons in the cortical area V4 are less understood compared to early visual areas V1 and V2. This is primarily because of V4 neurons’ highly nonlinear response properties. As a consequence, building predictive models for these neurons has been one of the challenging tasks in computational neuroscience. Recently, models based on convolutional neural networks (CNNs) have shown promise in predicting the activity of V4 neurons. More importantly, interpreting CNN-based models has offered tools to understand V4 neurons’ functional properties through visualizing their pattern selectivity. These interpretations, however, are based on models with hundreds of convolutional filters. Therefore, it is challenging to present a sparse set of filter bases to model each V4 neuron. To address this limitation, we propose two algorithms to remove redundant filters in the CNN-based models of V4 neurons. First, CAR compression that prunes filters from the CNN based on the filter’s contribution to the image classification accuracy. CAR is a greedy compression scheme to obtain smaller and more interpretable CNNs, while achieving close to original accuracy. Second, RAR compression that prunes filters based on their contribution to the neural response prediction accuracy. Both CAR and RAR provide a new set of simpler accurate models for V4 neurons. These models achieve almost similar (for CAR) or higher (for RAR) accuracy compared to the original model. Using the compressed models, we are able to find a sparse set of filters that accurately model V4 neurons. We identify and visualize a total of 25 filters in AlexNet that accurately model V4 neurons in non-human primates. The features extracted by these 25 filters can predict the spike rates of 71 V4 neurons with an average correlation coefficient of 51%. By visualizing the patterns selected by these models, we further demonstrate that V4 neurons are modeled via curvature or texture filters, as well as other more complicated filters. Our results present one of the first efforts to bridge between large-scale convolutional models of neurons and interpretable sparse networks.

avatar for Reza Abbasi-Asl

Reza Abbasi-Asl

Departments of Neurology, Bioengineering and Therapeutic Sciences, UCSF

Monday July 20, 2020 7:00pm - 8:00pm CEST
Slot 15