Hi. I'm Trapit.

A third year PhD student with Prof. Andrew McCallum at UMass. I am broadly interested in deep learning, reinforcement learning and applications to text data, knowledge bases, and multi-agent systems.

Learn more about what I do

Research Interests

I am a third year PhD student in IESL working with Prof. Andrew McCallum. I am interested in deep learning and reinforcement learning. My current research is on knowledge representation and reasoning. Check out some of my publications and try out some of the related codes.

Work Experience

Summer 2017: I interned at OpenAI where I had the most amazing experience working on reinforcement learning
Summer 2016: I had a great time interning at Facebook in the Applied Machine Learning group where I worked on deep learning models for some NLP tasks
Before PhD: I worked with with Prof. Ravi Kannan and Prof. Chiranjib Bhattacharyya at Indian Institute of Science, on Bayesian models and approximate inference

Research Highlights

Competitive Self-Play: In our recent work with OpenAI, we found that self-play allows simulated AI to discover remarkable physical skills without explicitly designing rewards for such skills.
Check out the blog post.

Publications

Click the titles for more information about the work and to download paper/supplementary/code/data.

  • [New] Emergent Complexity via Multi-Agent Competition. Trapit Bansal, Jakub Pachocki, Szymon Sidor, Ilya Sutskever, Igor Mordatch. ArXiv, 2017.
  • [New] Continuous Adaptation via Meta-Learning in Nonstationary and Competitive Environments. Maruan Al-Shedivat, Trapit Bansal, Yuri Burda, Ilya Sutskever, Igor Mordatch, Pieter Abbeel. ArXiv, 2017.
  • [New] RelNet: End-to-end Modeling of Entities and Relations. Trapit Bansal, Arvind Neelakantan, Andrew McCallum. In NIPS Workshop on Automated Knowledge Base Construction (AKBC), 2017.
  • Low-Rank Hidden State Embeddings for Viterbi Sequence Labeling. Dung Thai, Shikhar Murty, Trapit Bansal, Luke Vilnis, David Belanger, Andrew McCallum . In ICML Workshop on Deep Structured Prediction, 2017.
  • Ask the GRU: Multi-task Learning for Deep Text Recommendations. Trapit Bansal, David Belanger, Andrew McCallum. In ACM international conference on Recommender Systems (RecSys), 2016.
  • Content Driven User Profiling for Comment-Worthy Recommendations of News and Blog Articles. Trapit Bansal, Mrinal Das, Chiranjib Bhattacharyya. In ACM international conference on Recommender Systems (RecSys), 2015.
  • Ordered Stick-Breaking Prior for Sequential MCMC Inference of Bayesian Nonparametric Models. Mrinal Das, Trapit Bansal, Chiranjib Bhattacharyya. In International Conference on Machine Learning (ICML), 2015.
  • Relating Romanized Comments to News Articles by Inferring Multi-glyphic Topical Correspondence. Goutham Tholpadi, Mrinal Das, Trapit Bansal, Chiranjib Bhattacharyya. In AAAI conference on artificial intelligence (AAAI), 2015.
  • A Provable SVD-based Algorithm for Learning Topics in Dominant Admixture Corpus". Trapit Bansal, Chiranjib Bhattacharyya, Ravindran Kannan. In Neural Information Processing Systems (NIPS), 2014.
  • Going Beyond Corr-LDA for Detecting Specific Comments on News & Blogs. Mrinal Das, Trapit Bansal, Chiranjib Bhattacharyya. In ACM international conference on Web Search and Data Mining (WSDM), 2014.

Other Fun Stuff

  • Character LSTM for Sentiment Analysis on Twitter: Kate Silverstein, Jun Wang and I worked on a character-level LSTM model for sentiment analysis on Twitter. Read our short report on some of the cool results from the model. Code is available on Github. Kate also has an interesting post on it here.
  • Learning to play Atari: For my Deep Learning class mid-term project, I tried to implement double Q-learning on Atari Games. Check out the videos of the learned agent playing Boxing and Pong (spoiler: they kick ass!).

Code

  • Competitive Multi-agent Environments: code for the environements in paper "Emergent Complexity from Multi-agent Competition".
  • Specific Correspondence Topic Models (SCTM): C code for the WSDM'14 paper. Implements collapsed Gibbs sampling for three models: LDA, CorrLDA and SCTM. This is one of the few implementations I know for CorrLDA on text data. Also supports the feature of "sparse" topic distributions for LDA and CorrLDA.
  • Thresholded SVD (TSVD): Matlab code for the NIPS'14 paper. Implements the Thresholded SVD based K-means algorithm for topic recovery. This is much faster than Gibbs sampling and works easily for upto 200,000 documents (more if you have the RAM for it). Give it a spin.

Contact Me ...

Best way is to email me at "trapitbansal at gmail dot com" and I'll get back to you.