Cognitive CNN: Mimicking Human Cognition to resolve Shape-Texture bias

Recent works have demonstrated that CNNs trained on ImageNet learn to base their predictions only on the texture of the object in images, neglecting the information about the shape of the object. This is referred to as the texture bias hypothesis. We show that this bias is dependent on the dataset used to train the network, and propose a new regularizer based on attention matching to quantify and reduce this bias.
The work appeared at BAICS workshop at ICLR 2020.

What if Neural Networks had SVDs?

In this work we develop a highly parallelizable algorithm for multiplication of orthogonal matrices using the Householder parameterization. We achieved large improvements in running time compared to previous approaches. We also implemented our algorithm in CUDA and packaged it as a PyTorch submodule for easy usage.
The work appeared as a Spotlight paper at NeurIPS 2020.

MIMOQA:Multimodal Input Multimodal Output Question Answering

In this work we propose a new task for information retrieval in documents with multiple modalities. We also present a new dataset for the task of question answering with multiple modalities. Finally, we come up with a transformer based architecture to harness cross-modal attention and demonstrate its effectiveness for this task.
The work will appear at NAACL 2021.

Rule Augmented Unsupervised Constituency Parsing

Unsupervised parsing of syntactic trees has gained considerable attention. Aprototypical approach to such unsupervised parsing employs reinforcement learning and auto-encoders. However, no mechanism ensures that the learnt model leverages the well-understood language grammar. We propose an approach that utilizes very generic lingustic knowledge of the language present in the form of syntactic rules, thus inducing better syntactic structures. We introduce a novel formulation that takes advantage of the syntactic grammar rules and is independent of the base system.
The work will appear in the Findings of ACL 2021.