Misha

Misha Bilenko


misha@bilenko.com

Research   |   Personal


Update: I work at Meta on making large models useful and delightful to users.

Before that, I was at Google DeepMind working on Gemini assistant and shipping large models in products.

Before that, I was (back) at Microsoft working on deploying, applying and creating large ML models. I worked on Phi-3 models, Azure AI Services (Speech, Vision, Responsible AI, FormRecognizer, Language and Machine Translation), GitHub Copilot, Bing Chat, Azure OpenAI - as well as a few other Microsoft products, and with many amazing people.

Before that, I lived in Moscow and led the Machine Intelligence and Research (MIR) division in Yandex. We had all sorts of serious fun, including but not limited to: intelligent assistants (go Alice!), smart speakers (go Station!), speech recognition and synthesis, computer vision, machine translation, machine learning algorithms and platforms, and research in all of the above (go numerous Yandex products with all of the above!). Watching Speech, MT and Assistants move from statistical ML stacks to DNNs and ship lots of features and quality gains was really cool, launching Alice on all surfaces and several Station models was even better.

Before that, I lived in Seattle area and led the Machine Learning Algorithms team in Cloud+Enterprise division at Microsoft. Our ML tools were used in many products, from Microsoft AzureML, to SQL Server, to numerous others across all divisions of the company. We collaborated extensively with MSR and many applied ML/Data Science groups.

Before that, I was a researcher in the Machine Learning Department at Microsoft Research. I enjoyed building ML systems and tools, and working on large-scale prediction problems on behavioral, transactional and textual data. Specific applications on which I focused were high-throughput ML, click probability prediction, relevant ad selection, constructing user profiles for targeting, and improving search relevance with logs of user behavior. Earlier, I worked on semi-supervised clustering and record linkage (a.k.a. entity resolution a.k.a. de-duplication a.k.a. identity uncertainty a.k.a. co-reference resolution...).

Before that, I completed my Ph.D. in the Department of Computer Science at the University of Texas at Austin in 2006, where I was a member of the Machine Learning Group. Along the way, I spent the summer of 2002 at IBM T.J. Watson Research Center, and the summer/fall of 2004 at Google.

Besides all kinds of ML applications, I'm generally interested in adaptive similarity(distance, kernel, divergence, embedding, ...) functions, implementing learning algorithms on parallel/distributed platforms, and tools for machine learning practitioners. Evaluation methodology/metrics for all of these, offline and online, is always fun to understand and improve.

Research

  • Large-scale Learning
  • Learnable similarity functions and their applications in information integration (e.g., record linkage/identity uncertainty) and text mining


  • Semi-supervised clustering

    • Probabilistic Semi-Supervised Clustering with Constraints
      Sugato Basu, Mikhail Bilenko, Arindam Banerjee, and Raymond J. Mooney. In Semi-Supervised Learning, O. Chapelle, B. Sch�lkopf, and A. Zien (eds.), MIT Press, 2006.
      Note: this chapter summarizes the KDD and ICML papers below
      [PDF]

    • A Probabilistic Framework for Semi-Supervised Clustering
      Sugato Basu, Mikhail Bilenko, and Raymond J. Mooney. In Proceedings of the 10th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining (KDD-2004), pp.59-68, Seattle, WA, August 2004.
      (Winner of Best Research Paper Award)
      [PDF]

    • Integrating Constraints and Metric Learning in Semi-Supervised Clustering
      Mikhail Bilenko, Sugato Basu, and Raymond J. Mooney. In Proceedings of the 21st International Conference on Machine Learning (ICML-2004), pp.81-88, Banff, Canada, July 2004.
      [PDF]

    • A Comparison of Inference Techniques for Semi-supervised Clustering with Hidden Markov Random Fields
      Mikhail Bilenko and Sugato Basu. In Proceedings of the ICML-2004 Workshop on Statistical Relational Learning and its Connections to Other Fields (SRL-2004), pp.17-22, Banff, Canada, July 2004.
      [PDF]

  • Indirect learning in information integration (record linkage, information extraction), text classification, and clustering

    • Two Approaches to Handling Noisy Variation in Text Mining
      Un Yong Nahm, Mikhail Bilenko, and Raymond J. Mooney. In Proceedings of the ICML-2002 Workshop on Text Learning (TextML'2002), pp.18-27, Sydney, Australia, July 2002.
      [PDF]

Personal
In my leisure time I enjoy applying hill-climbing search and gradient descent algorithms to real-world domains, which are almost as cool as the cool stuff that my sister does.

Update: all outdoorsy hills-climbing searches and gradient descents are of the kid-friendly kind now. Much smaller step sizes and more regularized, still very fun!