Markos Athanasiadis: Attack-based identification of most-informative directions in binary classification

Zoom Lecture!

Abstract 

 Multivariate decoding that allows linking specific features of neuronal activity to cognitive functions and behavior can be implemented via machine learning algorithms for classification. Several reliable classifier implementations exist, such as SVMs and DNNs, that are capable of providing accurate predictions of  behavioral categorical variables from neuronal activity (fMRI, spike patterns). However the black-box nature of classifier based decoders prohibits understanding of the geometry of decision boundaries, which in turn hampers our ability to identify the neuronal activity features that are most decisive for a specific behavioral outcome.  

 Brute force methods for the visualization of the decision boundary are generally impractical owing to the high dimensionality of neural datasets. Recently, however,  methods have been developed that are capable of probing the decision boundary of classifiers by introducing minimal perturbations to the patterns that evoke misclassifications.  

 We develop a method to identify the most-informative directions in linear and non-linear, binary classification tasks making use of such adversarial perturbations. As a proof of principle, we validate our pattern identification method against the ground truth of artificially generated datasets. We observe that our method can follow closely, and in particular cases, outperform traditionally well-tested classifiers (linear-ANNs). Additionally, we attempt to decode population rate patterns from hippocampal activity during a spatial working memory task. We identify the most informative activity patterns during the delay period, and use them to extract single cell  information.  

 About the speaker and his research

 Hosted by Han Lu