{"id":2407,"date":"2025-02-20T12:39:16","date_gmt":"2025-02-20T12:39:16","guid":{"rendered":"https:\/\/meyn.ece.ufl.edu\/?page_id=2407"},"modified":"2026-03-17T12:02:43","modified_gmt":"2026-03-17T17:02:43","slug":"on-the-statistical-complexity-and-instance-adaptiveness-of-offline-reinforcement-learning","status":"publish","type":"page","link":"https:\/\/faculty.eng.ufl.edu\/meyn\/c3\/c3-9\/on-the-statistical-complexity-and-instance-adaptiveness-of-offline-reinforcement-learning\/","title":{"rendered":"What can we learn from signals and systems in a transformer?"},"content":{"rendered":"\n

Prashant Mehta<\/a> (UIUC)<\/h3>\n\n\n\n

Talk Abstract:  <\/strong>The transformer is the core algorithm underlying large language models (LLMs). In the decoder-only architecture, a finite sequence of tokens is mapped to a conditional distribution over the next token.<\/p>\n\n\n\n

In this talk, I place the transformer within the classical theory of prediction. In the 1940s, Wiener introduced the linear predictor, in which the conditional expectation of future data is expressed as a linear function of the past. I argue that the decoder-only transformer can be interpreted as a causal nonlinear generalization of this idea. Specifically, a transformer may be viewed as implementing a nonlinear predictor that approximates conditional distributions in sequence models.<\/p>\n\n\n\n

We develop this perspective in the setting of discrete-valued observations generated by a hidden Markov model (HMM), where connections to nonlinear filtering theory can be made explicit. The broader objective of this ongoing work is to bridge classical nonlinear filtering with modern transformer-based inference architectures. Joint work with Heng-Sheng Chang and Jin Won Kim. See https:\/\/www.arxiv.org\/abs\/2508.20211<\/a>.<\/p>\n\n\n

\n
\"\"<\/figure>\n<\/div>\n\n\n

Bio: Prashant Mehta<\/strong> received his M.S. degree in Electrical and Computer Engineering from the University of Massachusetts at Amherst in 1996 and the Ph.d. degree in Applied Mathematics from Cornell University in 2004 (with Prof. Tim Healey). He is now an Associate Professor of Mechanical Science and Engineering at the University of Illinois at Urbana-Champaign (UIUC). He is affiliated with the Coordinated Science Laboratory, the Department of Electrical and Computer Engineering, and the Neuroscience program at UIUC. He was the co-founder and the Chief Science Officer of the startup Rithmio whose gesture recognition technology was acquired by Bosch Sensortec in 2017. Prior to his academic appointment at UIUC in 2005, he worked at United Technologies Research Center (1996-98 and 2002-05) where he invented the symmetry-breaking solution to suppress combustion instabilities. This solution \u2014 which helped solve a sixty-year-old open problem \u2014 has since become an industry standard and is widely deployed in jet engines and afterburners sold by Pratt & Whitney.<\/p>\n\n\n\n

At UIUC, Prashant works with his students on mathematical and computational aspects of dynamics and control theory. His most significant recent contribution is the feedback particle filter algorithm (see recent projects for details on ongoing research in his lab). He is passionate about working with and mentoring undergraduate and graduate students. His students have won the best paper awards at IEEE CDC 2007 and 2009 and have been finalists for the best paper awards at ACC 2010 and at IEEE CDC 2012. He teaches dynamics and control classes as well as the graduate course sequence on engineering mathematics (see list of courses taught). For his teaching, he frequently makes the Incomplete list of teachers rated as excellent at UIUC.<\/p>\n\n\n\n

He has served as the Associate Editor for the IEEE Transactions on Automatic Control (2019-), Systems and Control Letters (2011-14) and the ASME Journal of Dynamic Systems, Measurement and Control (2012-16). He guest-edited the special issue of the ASME Journal of Dynamic Systems, Measurement and Control to commemorate the life, achievements and impact of Rudolph E. Kalman.<\/p>\n\n\n\n

<\/p>\n\n\n\n

  <\/p>\n\n\n","protected":false},"excerpt":{"rendered":"

Prashant Mehta (UIUC) Talk Abstract:  The transformer is the core algorithm underlying large language models (LLMs). In the decoder-only architecture, a finite sequence of tokens is mapped to a conditional distribution over the next token. In this talk, I place the transformer within the classical theory of prediction. In the 1940s, Wiener introduced the linear […]<\/p>\n","protected":false},"author":1347,"featured_media":0,"parent":2631,"menu_order":5,"comment_status":"closed","ping_status":"closed","template":"page-templates\/page-section-nav.php","meta":{"_acf_changed":false,"inline_featured_image":false,"featured_post":"","footnotes":"","_links_to":"","_links_to_target":""},"class_list":["post-2407","page","type-page","status-publish","hentry"],"acf":[],"_links":{"self":[{"href":"https:\/\/faculty.eng.ufl.edu\/meyn\/wp-json\/wp\/v2\/pages\/2407","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/faculty.eng.ufl.edu\/meyn\/wp-json\/wp\/v2\/pages"}],"about":[{"href":"https:\/\/faculty.eng.ufl.edu\/meyn\/wp-json\/wp\/v2\/types\/page"}],"author":[{"embeddable":true,"href":"https:\/\/faculty.eng.ufl.edu\/meyn\/wp-json\/wp\/v2\/users\/1347"}],"replies":[{"embeddable":true,"href":"https:\/\/faculty.eng.ufl.edu\/meyn\/wp-json\/wp\/v2\/comments?post=2407"}],"version-history":[{"count":4,"href":"https:\/\/faculty.eng.ufl.edu\/meyn\/wp-json\/wp\/v2\/pages\/2407\/revisions"}],"predecessor-version":[{"id":2705,"href":"https:\/\/faculty.eng.ufl.edu\/meyn\/wp-json\/wp\/v2\/pages\/2407\/revisions\/2705"}],"up":[{"embeddable":true,"href":"https:\/\/faculty.eng.ufl.edu\/meyn\/wp-json\/wp\/v2\/pages\/2631"}],"wp:attachment":[{"href":"https:\/\/faculty.eng.ufl.edu\/meyn\/wp-json\/wp\/v2\/media?parent=2407"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}