I am broadly interested in Natural Language Processing, Speech Recognition and Machine Learning Security & Privacy.

Research Papers

Thieves on Sesame Street! Model Extraction of BERT-based APIs
Kalpesh Krishna, Gaurav Singh Tomar, Ankur P. Parikh, Nicolas Papernot, Mohit Iyyer
ICLR 2020
blog // project page (code + slides + Twitter + external coverage)

Generating Question-Answer Hierarchies
Kalpesh Krishna, Mohit Iyyer
ACL 2019
project page (demo + code + data) // technical note // poster // external blog

Syntactically Supervised Transformers for Faster Neural Machine Translation
Nader Akoury, Kalpesh Krishna, Mohit Iyyer
ACL 2019
code // poster

Trick or TReAT: Thematic Reinforcement for Artistic Typography
Purva Tendulkar, Kalpesh Krishna, Ramprasaath R. Selvaraju, Devi Parikh
ICCC 2019 (oral presentation, Best Presentation Award)
code // slides // video // demo

Revisiting the Importance of Encoding Logic Rules in Sentiment Classification
Kalpesh Krishna, Preethi Jyothi, Mohit Iyyer
EMNLP 2018 (oral presentation, short paper)
code + data // slides // video

Hierarchical Multitask Learning for CTC-based Speech Recognition
Kalpesh Krishna, Shubham Toshniwal, Karen Livescu
[external video]

A Study of All-Convolutional Encoders for Connectionist Temporal Classification
Kalpesh Krishna, Liang Lu, Kevin Gimpel, Karen Livescu
ICASSP 2018 (Awarded SPS Travel Grant)
[poster]

Collaborators (in order of publication date): Karen Livescu, Kevin Gimpel, Liang Lu, Shubham Toshniwal, Preethi Jyothi, Mohit Iyyer, Purva Tendulkar, Ramprasaath R. Selvaraju, Devi Parikh, Nader Akoury, Gaurav Singh Tomar, Ankur P. Parikh, Nicolas Papernot

Software Papers

SunPy: A Python package for Solar Physics
Stuart J. Mumford and others
JOSS 2020

Thesis

Undergraduate Thesis - Constraint Driven Learning
(under guidance of Prof. Preethi Jyothi)
IIT Bombay (2017-2018)

Course Materials

Homework on linguistic probe tasks designed for UMass Amherst’s grad NLP class using AllenNLP.

Other Research (Course Projects)

MixMatch on Vision + Language Tasks (NLVR2): An attempt to integrate the MixMatch data augmentation algorithm for semi-supervised image classification to the challenging setting of NLVR2, where the input space has both images and text (report).

Research Exchange - A Collaborative Paper Annotation Tool - A platform to collaboratively annotate scientific literature to help newcomers understand research papers, built during an Human Computer Interaction course project (report).

Inference Networks for Structured Prediction - A TensorFlow implementation for the multi-label classification experiments in Learning Approximate Inference Networks for Structured Prediction. Also contains experiments on the FIGMENT dataset and a extension to Inference Network training algorithm based on Wasserstein GANs (report).

Diversity Sampling in Machine Learning - An implementation of Diverse Beam Search for Neural Networks in Language Modelling. Also contains the original (slightly modified code) for the interactive segmentation experiments in Diverse M-Best Solutions in MRFs (report).

Macro Actions in Reinforcement Learning - A suite of five algorithms (including ideas from “Learning to Repeat: Fine Grained Action Repetition for Deep Reinforcement Learning”) encouraging agents to repeat actions (report).

Single Image Haze Removal - An implementation of He et al. 2009, “Single Image Haze Removal using Dark Channel Prior” and an ongoing implementation of Bahat & Irani 2016, “Blind Dehazing using Internal Patch Recurrence” (report).

CNNs for Sentence Classification - A TensorFlow 1.1 implementation of Kim 2014, “Convolutional Neural Networks for Sentence Classification”.

Brittle Fracture Simulation - Python implementation of O’Brien and Hodgins 1999, “Graphical Modeling and Animation of Brittle Fracture”.

ECG Signal Analysis - Python implementation of parts of Christopher Buck, Aneesh Sampath 2013, “ECG Signal Analysis for Myocardial Infarction Detection.”.

Indian Language Datasets

As a part of my RnD project at IIT Bombay, I am releasing the dataset used to train my neural network language models. These have been mined from Wikipedia and I hope this will help further research in language modelling for Indian morphologically rich languages. The folder also contains the original PTB dataset.

  • Malayalam (denoted by ml)
  • Tamil (denoted by ta)
  • Kannada (denoted by kn)
  • Telugu (denoted by te)
  • Hindi (denoted by hi)
  • PTB (denoted by ptb)

All these datasets are compatible with SRILM. Files marked with unk have replaced all singletons with <unk> tokens. Files marked with char are character versions. All datasets have a train, valid and test file. You will find the dataset here.