An important by-product of Leonid's research is an efficient and flexible library for k-NN search codenamed NMSLIB created in collaboration with several other folks. Thanks to the contribution of Yury Malkov, the library was adopted by Amazon. The core retrieval method HNSW contributed by Yury was also reimplemented in the Facebook library FAISS. A brief description of this collaboration can be found on this LTI news page. This work is discussed in a podcast with Radim Řehůřek (author of Gensim) in March 2018. Feel free to check our code on GitHub.

NMSLIB integrates with another retrieval toolkit called FlexNeuART, which in August, November and December 2020 produced best neural and traditional submissions on the MS MARCO document ranking leaderboard. In that, the strongest traditional run outperformed a number of neural systems.

Leonid also co-authored an extremely efficient algorithm for light-weight compression of sorted integer numbers. We show that this algorithm can decompress at the speed of reading from memory. You can find software on GitHub. This software grew out of a now-popular library FastPFor. FastPFor has Python bindings.

Leonid was a graduate research assistant (aka PhD student) at the Language Technologies Institute at Carnegie Mellon University (under the supervision of Professor Eric Nyberg). In his thesis "Efficient and Accurate Non-Metric k-NN Search with Applications to Text Matching" he explored how various linguistic, neural, and lexical features can be incorporated directly into a candidate generation component (via k-NN search). He was assisting in teaching the following courses & seminars: Algorithms for NLP (11-711) in 2013, Software Engineering I (11-791) and Data Science Seminar (11-631) in 2014.

Featured blog posts:

• Understanding Performance of Long-Document Ranking Models through Comprehensive Evaluation and Leaderboarding.
• A Systematic Evaluation of Transfer Learning and Pseudo-labeling with BERT-based Ranking Models.
• Exploring Classic and Neural Lexical Translation Models for Information Retrieval: Interpretability, Effectiveness, and Efficiency Benefits.
• Traditional IR rivals neural models on the MS MARCO Document Ranking Leaderboard.
• A brief overview of classic and adversarial data augmentation techniques for speech recognition.
• Bringing a large Russian QA data set to light.
• Universal Methods to Sort Complex Information Tested.
• Efficient and Accurate Non-Metric k-NN Search with Applications to Text Matching: We Need More k-NN Search!
• Demystifying IBM Watson.
• Early life of dynamic programming.
• Will natural language processing engineers find it hard to get work in the future (once computers are capable of near-perfect text and speech processing)?
• Does Locality Sensitive Hashing (LSH) analysis have a fatal flaw?
• My awesome cat doing dog tricks (also see an explanatory blog post).

Additional information:

• Blog / Quora / Twitter / LinkedIn
• Publications/Talks (when applicable with software & blog links) (Google Scholar)
• Software (GitHub page)
• Datasets