Kumar Krishna Agrawal

I am first year PhD student at Berkeley AI Research (BAIR), UC Berkeley.

Previously, I was a researcher at Google Brain, as part of the Residency program. My research is motivated by the need to design algorithms which are resource (data / compute) efficient and theoretically grounded. Broadly, I am interested in machine learning, robotics and languages (programming / natural). I enjoy bringing together insights from fundamental research and algorithm design, to build sytems which work in the real-world.

Previously, I graduated from Indian Institute of Technology Kharagpur majoring in Mathematics and Computing. In the past, I've been fortunate to work under the guidance of Prof. Yoshua Bengio, Prof. Raman Arora and Prof. B. Sury.

blog / github / gscholar

Kumar Krishna Agrawal

I am broadly interested in foundations of machine learning and optimization, with a focus on developing efficient algorithms for robotics. More specifically, I am interested in enhanced perception, control and reasoning, to allow robots to replicate and generate complex behaviour.

GANSynth: Adversarial Neural Audio Synthesis
Jesse Engel, Kumar Krishna Agrawal, Shuo Chen, Ishaan Gulrajani,
Chris Donahue, Adam Roberts
International Conference on Learning Representations (ICLR), 2019
arXiv / samples

We demonstrate that GANs are able to outperform strong WaveNet baselines on automated and human evaluation metrics, and efficiently generate audio ~54,000 times faster than their autoregressive counterparts.

Discriminator Actor Critic: Addressing sample inefficiency and reward bias in Adversarial Imitation Learning
Ilya Kostrikov, Kumar Krishna Agrawal, Debidatta Dwibedi,
Sergey Levine, Jonathan Tompson
International Conference on Learning Representations (ICLR), 2019

We identify and propose an approach to address reward bias in Adversarial Imitation Learning Algorithms. Combining with off-policy learning we demonstrate order of magnitude improvement in sample complexity for this class of algorithms.

Towards Mixed Optimization for Reinforcement Learning with Program Synthesis
Surya Bhupatiraju*, Kumar Krishna Agrawal*, Rishabh Singh
Workshop on Neural Abstract Machines and Program Induction, ICML 2018

Playing with Embeddings : Evaluating embeddings for Robot Language Learning through MUD Games
Anmol Gulati*, Kumar Krishna Agrawal*
Workshop on Evaluating Vector Space Representations, EMNLP 2017
arXiv / slides

Recurrent Memory Addressing for describing videos
Arnav Jain*, Abhinav Agarwalla*, Kumar Krishna Agrawal*, Pabitra Mitra
Workshop on Deep Learning in Computer Vision, CVPR 2017

Generative Adversarial Motion to Image Synthesis
Arna Ghosh*, Kumar Krishna Agrawal*
Workshop on Artificial and Biological Vision, ECCV 2016

(This, is much more classy)