Скачать с ютуб Lecture 8, 2025; GPT, HMM, and Markov chains: Rollout variants for most likely sequence generation в хорошем качестве

Lecture 8, 2025; GPT, HMM, and Markov chains: Rollout variants for most likely sequence generation

Slides, class notes, and related textbook material at https://web.mit.edu/dimitrib/www/RLbo... Lecture delivered by Dimitri Bertsekas and Yuchao Li, based on the recent arXiv paper: "Most Likely Sequence Generation for n-Grams, Transformers, HMMs, and Markov Chains, by Using Rollout Algorithms," by Y. Li and D. Bertsekas, arXiv:2403.15465, March, 2024. Slides can be found at https://web.mit.edu/dimitrib/www/RLTo... Most Likely Generated Sequences in $n$-Grams; Inference in Hidden Markov Models (HMM) contexts, such as those relating to the use of the Viterbi algorithm; Applications in speech recognition, language translation, computational linguistics, coding and error correction, bioinformatics, investment and trading, etc; DP Formulation of most likely sequence selection problem; Rollout Algorithms and performance improvement; Double rollout. Computational experiments with Markov chains; Computational experiments with Large language Models and a fine-tuned GPT trained with gradient descent/Adam, and data from the works of the American author J. Michener