Files related to my intern project at Microsoft Research New England where I am investigating the interpretability of CNNs for medical image classification.

This work is based on the "network dissection" technique developed in Antonio Torralba's lab by David Bau, Bolei Zhou, et al - http://netdissect.csail.mit.edu/final-network-dissection.pdf. This technique aligns network units with visual concepts that humans are familiar with in order to interpret how convolutional neural networks learn images.

The NetDissect folder contains the implementation code that accompanied their paper, along with some changes that I made in order to make the technique applicable to the datasets I was using in addition to the dataset they used to obtain their original results. I issued a pull request for two of those changes which were accepted by David Bau.