Predicting Retinal Ganglion Cell Responses Based on Visual Features via Graph Filters

This thesis presents a novel graph-based approach to classify video clips with binary labels. Each video clip is described by a feature vector instead of raw pixel values.

At the model's core, a similarity graph is defined where each node is associated with a feature vector and its corresponding label. The weight of an edge connecting two nodes delineates the similarity of the nodes' feature vectors which is computed via the Mahalanobis distance and its metric matrix. The metric matrix is learned using labeled training data. Unknown labels are then estimated using the optimized metric and the similarity graph.

The main advantage of our model is enabling interpretations of how the predictions are made. Nevertheless, the model achieves competitive accuracy with state-of-the-art approaches as well. We apply this model to a retinal coding problem where explainability is essential to gain conceptual insight about the retina.