Koustuv Sinha

Koustuv Sinha

Research Scientist

Meta AI

My current research interest involves investigating the limits of systematic language understanding of modern neural language representation systems, by leveraging linguistic and logical priors. In particular, I investigate how language is represented by neural models in a human-like way by testing their ability to understand the semantics, syntax and generalizability in the context of natural and artificial languages. I am also interested in improving the state of Generative Language Modeling (Dialog Systems).

I’m a Research Scientist at Meta AI New York, in the Fundamental AI Research (FAIR) team. I did my PhD from McGill University (School of Computer Science) and Mila (Quebec AI Institute), supervised by Joelle Pineau, in the wonderful city of Montreal, QC, Canada. I spent a significant portion of my PhD being a Research Intern (STE) at Meta AI (FAIR), Montreal.

I am an associate editor of ReScience C, a peer reviewed journal promoting reproducible research, and I am the lead organizer of the annual Machine Learning Reproducibility Challenge (V1, V2, V3, V4, V5 ). My work has been covered by several news outlets in the past, including Nature, VentureBeat, InfoQ, DailyMail and Hindustan Times.

Resumé | PhD Thesis

Interests
  • Machine Learning
  • Natural Language Processing
  • Computational Linguistics
Education

  • PhD in Computer Science (ML & NLP), 2022

    McGill University

  • MSc in Computer Science (ML & NLP), 2018

    McGill University

  • B.Tech in Computer Science, 2014

    West Bengal University of Technology

Recent Posts

LLMs can sanitize annotations! Using zero shot relation extraction to fix CLUTRR templates
Introduction It has been three years since the release of CLUTRR, a benchmark we created to test the reasoning capabilities of modern neural networks. The idea is simple: can models understand first-order logic, in the backdrop of kinship relations?
Dec 23, 2022 11 min read
Replicating Zotero-connector functionality in Emacs … without Zotero!
In my last blog post I described a method I use to keep track of my paper reading habits, using Emacs. Using the workflow, I can now: Check the latest Arxiv papers using Elfeed Score the papers using Elfeed-Score Save the papers in a local bib file, along with pdfs, using org-ref functions Maintain a paper reading tracker document in Org Mode, where the workflow automatically adds the paper to read.
Oct 12, 2022 7 min read
A workflow for reading, managing and discovering ML research papers with Emacs
Over the last couple of years I have steadily transferred most of my workflows in Emacs (more specifically, Doom Emacs). As they truly say, Emacs is not just an editor, it is an operating system.
Jul 18, 2022 11 min read
See all posts

Recent News

All news»

Recent Publications

Koustuv Sinha, Amirhossein Kazemnejad, Siva Reddy, Joelle Pineau, Dieuwke Hupkes, Adina Williams (2022). The Curious Case of Absolute Position Embeddings. Findings of Empirical Methods of Natural Language Processing.

Cite DOI Arxiv

Shanya Sharma, Manan Dey, Koustuv Sinha (2022). How sensitive are translation systems to extra contexts? Mitigating gender bias in Neural Machine Translation models through relevant contexts. Findings of Empirical Methods of Natural Language Processing (EMNLP), 2022.

Cite DOI Arxiv

Koustuv Sinha, Robin Jia, Dieuwke Hupkes, Joelle Pineau, Adina Williams, Douwe Kiela (2021). Masked Language Modeling and the Distributional Hypothesis: Order Word Matters Pre-training for Little. Empirical Methods of Natural Language Processing (EMNLP).

Cite Arxiv ACL Anthology Code

See all publications

Contact