Nik Tavakolian: Revealing the Inner Workings of Deep neural Networks: A Graph-Based Interpretability Framework

 Statistical classification is the process of predicting the category or label of data points based on observed features. For example, classifying the right animal from images of dogs and cats. Deep neural networks (DNNs) have surpassed traditional classification techniques in tasks like image and language classification. However, while DNNs achieve exceptional accuracy, their decision-making processes are mostly hidden, leaving questions about the specific computational steps that these models perform when classifying data.

In this talk, we explore how DNNs organize and transform data across their hidden layers, where intermediate representations of the input data are generated. We introduce a novel geometric framework, which can be applied to any data representation, enabling us to analyze how the data representations are changing throughout the layers of DNNs. We also present results of applying the framework on a ResNet-18 model, a modern DNN architecture, trained to perform classification on the MNIST and CIFAR-10 image datasets.