English

PLPP: Prompt Learning with Perplexity Is Self-Distillation for Vision-Language Models

Computation and Language 2024-12-23 v1 Artificial Intelligence

Abstract

Pre-trained Vision-Language (VL) models such as CLIP have demonstrated their excellent performance across numerous downstream tasks. A recent method, Context Optimization (CoOp), further improves the performance of VL models on downstream tasks by introducing prompt learning. CoOp optimizes a set of learnable vectors, aka prompt, and freezes the whole CLIP model. However, relying solely on CLIP loss to fine-tune prompts can lead to models that are prone to overfitting on downstream task. To address this issue, we propose a plug-in prompt-regularization method called PLPP (Prompt Learning with PerPlexity), which use perplexity loss to regularize prompt learning. PLPP designs a two-step operation to compute the perplexity for prompts: (a) calculating cosine similarity between the weight of the embedding layer and prompts to get labels, (b) introducing a language model (LM) head that requires no training behind text encoder to output word probability distribution. Meanwhile, we unveil that the essence of PLPP is inherently a form of self-distillation. To further prevent overfitting as well as to reduce the additional computation introduced by PLPP, we turn the hard label to soft label and choose top-kk values for calculating the perplexity loss. For accelerating model convergence, we introduce mutual self-distillation learning, that is perplexity and inverted perplexity loss. The experiments conducted on four classification tasks indicate that PLPP exhibits superior performance compared to existing methods.

Keywords

Cite

@article{arxiv.2412.15277,
  title  = {PLPP: Prompt Learning with Perplexity Is Self-Distillation for Vision-Language Models},
  author = {Biao Liu and Wenyi Fang and Xiaoyu Wu and Yang Zheng and Zheng Hu and Bo Yuan},
  journal= {arXiv preprint arXiv:2412.15277},
  year   = {2024}
}
R2 v1 2026-06-28T20:42:54.795Z