Скачать или смотреть Tensor Networks, Probabilistic Modeling, and Formal Grammar with Jacob Miller

Abstract from Jacob:


Tensor networks (TNs) are a general formalism for efficiently modeling complex higher-order tensors, whose broad applications within quantum many-body physics have led them to be described as the natural framework for modeling quantum states of matter. Spurred on by their independent rediscovery in applied mathematics, TNs have increasingly been used for machine learning, where they have unlocked a growing number of novel theoretical and practical results. In this talk, we discuss one such result, concerning the use of TN models for modeling probabilistic sequence data. We show how a simple recurrent TN model enables a novel algorithm for conditional sampling from collections of strings defined by formal grammars, a significant generalization of the autoregressive sampling of typical language models. By developing this unexpected link between quantum physics and language, we expect our results to contribute to the search for mathematically principled methods for unsupervised grammatical inference in NLP.

https://docs.google.com/presentation/...