Speaker: Alejandro Pozas-Kerstjens
Affiliation: Université de Genève
Date: Thursday, 13 March 2025 at 15:00
Location: Seminar Room, Serrano 113b
The idea of using tensor networks as machine learning architectures is now almost 10 years old. In this (almost) decade, it has been shown that these architectures showcase benefits in aspects such as privacy and explainability. However, the scarcity in the types of architectures that we can train may be a limiting factor for large-scale deployments. Which begs the question: in the age of LLMs with billions of parameters, does it make sense to study tensor networks? In this talk I will argue that yes, it does make sense. The reason for it is the recent appearance of techniques that allow to give tensor-network form of functions accessed in a black-box manner. I will, in particular, describe the technique that we introduced in https://www.arxiv.org/abs/2501.06300, called Tensor train via recursive sketching from samples. This algorithm is naturally suited for machine learning environments, since it uses the training dataset to create a tensor-network approximation of arbitrary models. I will showcase how this method brings the benefits of tensor-network architectures to other types of models, and how it opens an exciting possibility for interpreting neural-network quantum states.