An engineer is training a model on very long documents and observes that the attention mechanism is behaving erratically. The model's focus shifts dramatically between tokens from one training step to the next, leading to poor convergence. A closer look at the attention weight distributions reveals they are often extremely "peaky," with one or two tokens receiving nearly all the weight (e.g., weights like [0.01, 0.98, 0.01]), and the location of this peak changes unpredictably. Which of the foll
0
1
Tags
Ch.2 Generative Models - Foundations of Large Language Models
Foundations of Large Language Models
Foundations of Large Language Models Course
Computing Sciences
Analysis in Bloom's Taxonomy
Cognitive Psychology
Psychology
Social Science
Empirical Science
Science
Related
An engineer is training a model on very long documents and observes that the attention mechanism is behaving erratically. The model's focus shifts dramatically between tokens from one training step to the next, leading to poor convergence. A closer look at the attention weight distributions reveals they are often extremely "peaky," with one or two tokens receiving nearly all the weight (e.g., weights like [0.01, 0.98, 0.01]), and the location of this peak changes unpredictably. Which of the foll
Mechanism of Attention Stabilization
The Role of Global Tokens in Mitigating 'Spiky' Attention