DistillCSE: Distilled Contrastive Learning for Sentence Embeddings
arXiv:2310.13499 · doi:10.18653/v1/2023.findings-emnlp.547
Abstract
This paper proposes the DistillCSE framework, which performs contrastive learning under the self-training paradigm with knowledge distillation. The potential advantage of DistillCSE is its self-enhancing feature: using a base model to provide additional supervision signals, a stronger model may be learned through knowledge distillation. However, the vanilla DistillCSE through the standard implementation of knowledge distillation only achieves marginal improvements due to severe overfitting. The further quantitative analyses demonstrate the reason that the standard knowledge distillation exhibits a relatively large variance of the teacher model's logits due to the essence of contrastive learning. To mitigate the issue induced by high variance, this paper accordingly proposed two simple yet effective solutions for knowledge distillation: a Group-P shuffling strategy as an implicit regularization and the averaging logits from multiple teacher components. Experiments on standard benchmarks demonstrate that the proposed DistillCSE outperforms many strong baseline methods and yields a new state-of-the-art performance.
References in corpus (10)
- Adam: A Method for Stochastic Optimization
- Efficient Estimation of Word Representations in Vector Space
- A Simple Framework for Contrastive Learning of Visual Representations
- Representation Learning with Contrastive Predictive Coding
- Unsupervised Representation Learning by Predicting Image Rotations
- Unsupervised Learning of Sentence Embeddings using Compositional n-Gram Features
- CLEAR: Contrastive Learning for Sentence Representation
- Whitening Sentence Representations for Better Semantics and Faster Retrieval
- Towards Understanding Ensemble, Knowledge Distillation and Self-Distillation in Deep Learning
- An Unsupervised Sentence Embedding Method by Mutual Information Maximization