On Transferability of Prompt Tuning for Natural Language Processing

1 minute read

Introduction

  • What is the name of the THUNLP prompt transfer paper?

    On Transferability of Prompt Tuning for Natural Language Processing

  • What are the main contributions of the THUNLP prompt transfer paper?
    • trained soft prompts can transfer zero shot to other tasks
    • can transfer across models with a cross model projector
    • as initialization soft prompts can significantly accelerate training
    • explore the effect of various prompt similarity metrics as transferability indicators
  • One issue with prompt tuning is it is sometimes slower to converge compared to full finetuning

    Untitled

  • roberta-large: zero shot transfer performance: best transfer prompt varies widely
    • do they use the same roberta classification head or a PET style verbalizer?

    Untitled

Method

  • Cross task and cross model soft prompt transfer

    Untitled

  • How can we transfer soft prompts across model architectures?
    • use a learnable projector (2 layer perceptron)
  • What objective can we use to train a cross model soft prompt projector?
    • distance minimization: minimize distance between prompt trained on the same task
    • task specific: backpropagate weights to the source prompt
      • hopefully can generalize to transfer prompts for other tasks

Results

  • THUNLP soft prompt transfer results: initialization transfer leads to marginal boost in performance
    • convergence speedup is quite high for T5

    Untitled

  • Cross Model prompt transfer results:
    • doesn’t seem to work particulary well

    Untitled

Conclusions

Similar in idea to the SPOT paper. I didn’t find the transferability indicators that interesting since a similar prompt is always likely to give the best transfer task and also the overlapping neuron activation rate was rather poorly motivated.

Reference

@article{su2021transferability,
  title={On transferability of prompt tuning for natural language understanding},
  author={Su, Yusheng and Wang, Xiaozhi and Qin, Yujia and Chan, Chi-Min and Lin, Yankai and Liu, Zhiyuan and Li, Peng and Li, Juanzi and Hou, Lei and Sun, Maosong and others},
  journal={arXiv preprint arXiv:2111.06719},
  year={2021}
}

Updated: