Research

Why and when should you pool?
Analyzing Pooling in Recurrent Architectures

Prof. Mausam, IIT Delhi

  • We show why and how Pooling (and attention) based BiLSTMs demonstrate improved learning ability and positional invariance over standard BiLSTMs.
  • Further, we find that Pooling helps improve sample efficiency (low-resource settings) and is particularly beneficial when important words lie towards the middle of the sentence
  • Our proposed pooling technique max-attention (MaxAtt) helps improve upon past approaches on standard accuracy metrics, and is more robust to distribution shift
  • Analyses done on multiple Text Classification tasks.

Adversarial Robustness against the Union of Multiple Perturbation Models

Carnegie Mellon University, Pittsburgh, USA

Prof. Zico Kolter, 2019

  • Designed an algorithm for robustness against the union of multiple perturbations types (L1, L2, Linf).
  • Showed improved robustness exceeding previous benchmarks on MNIST and CIFAR-10 datasets.
  • Accepted at ICML 2020.

The code and trained models for this work can be found on the Github link.

A manuscript for the same is available on ArXiv

Image Dehazing for Low Computation Devices

Samsung Research and Development Headquarters, Suwon, South Korea

Summer 2019

  • Worked on the task of image dehazing for low computation scenarios like mobile devices and refrigerators.
  • Designed a novel loss function to train a Generative Adversarial Network (GAN) for low resource settings.
  • Model was 1500 times faster & 200 times lighter than the state-of-the-art with acceptable loss in quality.
  • Received Pre-Placement Offer for exceptional performance. Work done is in preparation for submission.

PERSONA

École Polytechnique Fédérale de Lausanne, Switzerland

Prof. James Larus, Summer 2018

  • Worked on a TensorFlow based Cluster-scale, high-throughput bioinformatics framework, PERSONA.
  • Implemented Striped Smith-Waterman Algorithm to make gene sequencing computationally feasible.
  • Succeeded in speeding-up the sub-process of protein alignment by over 6 times. [Presentation]