Colloquia - Mithün Paul Paneghat

Are Neural Networks Learning Spurious Hidden Patterns?

by Mithün Paul Panenghat

Abstract: While neural networks produce state-of-the-art performance in many natural language processing tasks, it is suspected that they generally learn from lexical information, which transfer poorly between domains. In this talk we explore our work where we investigate the importance that a model assigns to various aspects of  data while learning and making predictions. By inspecting the attention weights assigned by the model, we confirm that most of the weights are assigned to noun phrases. To mitigate this dependence on lexicalized information, we experiment with strategies of masking out lexical information. Our results show that, while performance on the in-domain dataset remains on par with that of the model trained on fully lexicalized data, it improves considerably when tested out of domain. This proves our hypothesis that neural networks are possibly learning spurious hidden patterns based on lexical information and not linguistic principles as believed. We also suggest a possible solution using preprocessing of data by delexicalization thus enabling transfer learning across domains.

Bio: Mithun is a 4th year PhD student in the Department of Computer Science, where he is advised by Professor Mihai Surdeanu. Mithun is broadly interested in natural language processing applications driven by neural networks. His current research focuses on fact verification with an emphasis on improving the robustness of fact verification approaches when transferred between different domains. Mithun earned his Bachelor degree from Birla Institute of Technology and Science, Pilani, India , with a double major in Engineering and Physics, along with his first Masters in Physics. He did his second masters in Computer Science at the University of Arizona. His other interests include network security and penetration testing.


10 a.m. Nov. 1, 2019