Stabilizing Attention with Global Tokens
A key benefit of incorporating global tokens is the stabilization of model performance, particularly when processing very long sequences. By allowing all tokens to attend to a common set of global tokens, the output distribution of the Softmax function in the attention calculation is smoothed. This prevents the attention weights from becoming too sparse or concentrated, leading to more stable training and inference.
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
Learn After
An engineer is training a model on extremely long text sequences and notices that the training process is unstable. Upon inspection, they find that the attention mechanism frequently assigns nearly all of its weight to a single, specific token, ignoring the rest of the context. This extreme concentration of attention scores is causing gradients to become erratic. Which of the following strategies directly addresses this problem by encouraging a less concentrated, more stable distribution of atte
A machine learning engineer is training a model on very long documents and observes that the attention mechanism is behaving unstably: for some parts of the text, the attention weights are extremely concentrated on a single word, while for others, they are so spread out they are effectively zero. After implementing a change, the engineer observes that the attention weights are now more smoothly distributed, leading to more consistent model performance. Which of the following changes best explain
Evaluating Attention Mechanism Stability