Transformers learn through gradual rank increase
This provides incremental insights into transformer training mechanisms for researchers in machine learning.
The paper tackles the problem of understanding learning dynamics in transformers by identifying that the difference between trained and initial weights progressively increases in rank, and it proves this occurs under specific assumptions while showing it can happen in practice without them.
We identify incremental learning dynamics in transformers, where the difference between trained and initial weights progressively increases in rank. We rigorously prove this occurs under the simplifying assumptions of diagonal weight matrices and small initialization. Our experiments support the theory and also show that phenomenon can occur in practice without the simplifying assumptions.