The Power of Scale for Parameter-Efficient Prompt Tuning
Summary
This paper proposes prompt-tuning, learning a set of soft-prompts in the latent space instead of using natural language.
Approach
Comparison of prompt-tuning vs traditional model-tuning
Model tuning ends up creating different versions of the model specialized for each task, while prompt-tuning maintains one base model and only learns the task specialized prompts.
The authors examine prompt-tuning by transforming a T5 model into a language model (instead of the original reconstruction task).
Prompt initialization strategies
- Random initialization
- Use pre-trained vocabulary
- (For classification) Use pre-trained vocabulary to encode the classification targets.
Domain shift
- The authors argue that because prompt-tuning does not modify the generalist weights of the language model but rather modify the representation of the input to be task specific, prompt-tuning may be more resilient to domain shift.
- This is tested by training/testing on OOD dataset groups.
Ensembling
- Prompt-tuning can be a cheap alternative for ensembling.
- By learning \(N\) sets of prompt tokens, one can simply construct a batch of \(N\) inputs with each prompt and compute the output in one forward pass
Findings
- Prompt tuning becomes more comparable to model-tuning as model size increases
- Prompt tuning does show more resilience to domain shift compared to model fine-tuning.
- The learned prompts end up in a semantically meaningful position in the model's vocab space.
- The nearest neighbors of each learned prompt often form a coherent cluster.
- When encoding the classification target, the final prompt still shows similarity to the encoded target values.
- This also shows better performance over single-prompt tuning.
- The ensemble approach yields better performance than single-prompt settings.