Towards Interpretable and Controllable Representation Learning
One of the keys to the empirical successes of deep neural networks in many domains, such as natural language processing and computer vision, is their ability to automatically extract salient features for downstream tasks via the end-to-end learning paradigm. The end-to-end learning paradigm, however, comes at the expense of model interpretability and controllability.
In this talk, I will first present our recent work on using probing methods to investigate and interpret two learning properties in deep neural networks: (i) laziness — information already memorized in a module will not be propagated into other modules; (ii) targetedness — information that is unnecessary for the ultimate objective will be filtered out from the internal layers. Second, I will present how to utilize these two properties to control the learned representations — learning representations to decouple global and local information from/for image generation. I will conclude by laying out future research directions towards interpretable and controllable representation learning by establishing theoretical framework to formally link various neural architectures with the learned representations.
Xuezhe Ma is a final year PhD student at Carnegie Mellon University, advised by Eduard Hovy.
Before that, he received his B.E and M.S from Shanghai Jiao Tong University. His research interests fall in areas of natural language processing and machine learning, particularly in deep learning and representation learning with applications to linguistic structured prediction and deep generative models. Xuezhe has interned at Allen Institute for Artificial Intelligence (AI2) and earned the AI2 Outstanding Intern award. His research has been recognized with outstanding paper award at ACL 2016 and best demo paper nomination at ACL 2019.