Dynamic metastability in the self-attention model

Fuente: arXiv
Saved in:
Bibliographic Details
Main Authors: Geshkovski, Borjan, Koubbi, Hugo, Polyanskiy, Yury, Rigollet, Philippe
Format: Preprint
Published: 2024
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866913539210346496
author Geshkovski, Borjan
Koubbi, Hugo
Polyanskiy, Yury
Rigollet, Philippe
author_facet Geshkovski, Borjan
Koubbi, Hugo
Polyanskiy, Yury
Rigollet, Philippe
contents We consider the self-attention model - an interacting particle system on the unit sphere, which serves as a toy model for Transformers, the deep neural network architecture behind the recent successes of large language models. We prove the appearance of dynamic metastability conjectured in [GLPR23] - although particles collapse to a single cluster in infinite time, they remain trapped near a configuration of several clusters for an exponentially long period of time. By leveraging a gradient flow interpretation of the system, we also connect our result to an overarching framework of slow motion of gradient flows proposed by Otto and Reznikoff [OR07] in the context of coarsening and the Allen-Cahn equation. We finally probe the dynamics beyond the exponentially long period of metastability, and illustrate that, under an appropriate time-rescaling, the energy reaches its global maximum in finite time and has a staircase profile, with trajectories manifesting saddle-to-saddle-like behavior, reminiscent of recent works in the analysis of training dynamics via gradient descent for two-layer neural networks.
format Preprint
id arxiv_https___arxiv_org_abs_2410_06833
institution arXiv
publishDate 2024
record_format arxiv
spellingShingle Dynamic metastability in the self-attention model
Geshkovski, Borjan
Koubbi, Hugo
Polyanskiy, Yury
Rigollet, Philippe
Machine Learning
Analysis of PDEs
Dynamical Systems
We consider the self-attention model - an interacting particle system on the unit sphere, which serves as a toy model for Transformers, the deep neural network architecture behind the recent successes of large language models. We prove the appearance of dynamic metastability conjectured in [GLPR23] - although particles collapse to a single cluster in infinite time, they remain trapped near a configuration of several clusters for an exponentially long period of time. By leveraging a gradient flow interpretation of the system, we also connect our result to an overarching framework of slow motion of gradient flows proposed by Otto and Reznikoff [OR07] in the context of coarsening and the Allen-Cahn equation. We finally probe the dynamics beyond the exponentially long period of metastability, and illustrate that, under an appropriate time-rescaling, the energy reaches its global maximum in finite time and has a staircase profile, with trajectories manifesting saddle-to-saddle-like behavior, reminiscent of recent works in the analysis of training dynamics via gradient descent for two-layer neural networks.
title Dynamic metastability in the self-attention model
topic Machine Learning
Analysis of PDEs
Dynamical Systems
url https://arxiv.org/abs/2410.06833