Training Dynamics of Transformers on Unbounded Inputs: Attention Learning Can Lag Output Layer Learning

Key points

• The success of transformers is attributed to several factors, including their capacity to express formal languages over variable-length inputs, and ability of the attention mechanism to focus on relevant input features.1
• A natural extension of the theory of one-step gradient feature learning in feedforward networks suggests that one could first learn the attention weights with a single gradient step, and then train the output layer.1
• We prove that the transformer can be trained to recognize the language with a three-phase gradient descent: the output layer is updated first, followed by an attention update, and the output layer is fine-tuned in the third phase.1
• We empirically show that "attention learning lagging output layer learning" is beneficial in other language recognition, regression, and in-context learning tasks as well.1

Sources (1)

• [1]Training Dynamics of Transformers on Unbounded Inputs: Attention Learning Can Lag Output Layer Learning
  arXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 14, 03:19 PM

  We show that this fails for a shallow transformer trained to recognize a star-free language from unbounded-length input.

  The success of transformers is attributed to several factors, including their capacity to express formal languages over variable-length inputs, and ability of the attention mechanism to focus on relevant input features.

Extractive summary: sentences quoted from the sources.