Publications

Learning Robust Neural Networks using Wasserstein Adversarial GAN (WAGAN)

Published in , 2021

At present, studies of adversarial examples have relied on Lp-norm to evaluate perceptual similarity. Recent studies, however, have found that Lp-norm is an insufficient and inadequate measure of perceptual distance between images. Our research proposes a novel approach to generate adversarial examples that have large Lp-norms, but are perceptually more similar to the original input and establish state-of-the-art results in robustness.

Recommended citation: Shashank Goel*, Parth Shah*, Harini Suresh /files/pdf/research/wagan.pdf

Interpreting BERT for downstream tasks: Empirical Study on Coreference Resolution

Published in , 2021

The study adapts existing methods of attribution to explain BERT’s coreference resolution decision. Furthermore, this work provides guidance for using attribution methods to analyze BERT’s decision-making in downstream tasks.

Recommended citation: Parth Shah, Jing Xu, Jieyu Zhao, Fan Yin /files/pdf/research/interpreting_bert.pdf

Regex Queries over Incomplete Knowledge Bases Permalink

Published in Automated Knowledge Base Construction (AKBC), 2021, 2021

We propose the novel task of answering regular expression queries (containing disjunction and Kleene plus operators) over incomplete KBs and solve it using RotatE-Box – a novel combination of RotatE and box embeddings. RotatE-Box is tested on two new regex-query datasets introduced in this paper, including one that is harvested from actual user query logs, and we find that it significantly outperforms models built on just RotatE and just box embeddings.

Recommended citation: Vaibhav Adlakha, Parth Shah, Srikanta Bedathur, Mausam, Published at Automated Knowledge Base Construction (AKBC), 2021 /files/pdf/research/regex_queries_over_incomplete_kb.pdf

^* Equal authorship