Deep neural networks, widely used for pattern recognition tasks such as vision classification problems, are often regarded as black boxes. For deep neural networks to be interpretable, efforts have been made to uncover and visualize the types of features learned by each neuron. Previous work by the lab has provided a Network Dissection framework, which uses Borden, a labeled dataset, to align each neuron’ s activations with human-interpretable concepts. I will use this framework and its interpretability metric to catalogue how various training methods affect a network’ s interpretability. Using this understanding, I hope to ultimately propose practical methods for deep learning that maximize interpretability while producing competitive results.

When I took Advances in Computer Vision (6.819) some semesters ago, I became interested in machine learning. After doing a few related projects in classes and internships, I am excited to now use this SuperUROP project as an opportunity to dive deeper and to be able to work on a problem I personally find unique and fulfilling.