S-Prompts Learning with Pre-trained Transformers: An Occam's Razor for Domain Incremental Learning

By Yabin Wang, Zhiwu Huang, Xiaopeng Hong. 2022

371 citations recorded by the source corpus.

State-of-the-art deep neural networks are still struggling to address the catastrophic forgetting problem in continual learning. In this paper, we propose one simple paradigm (named as S-Prompting) and two concrete approaches to highly reduce the forgetting degree in one of the most typical continual learning scenarios, i.e., domain increment learning (DIL). The key idea of the paradigm is to learn prompts independently across domains with pre-trained transformers, avoiding the use of exemplars that commonly appear in conventional methods. This results in a win-win game where the prompting can achieve the best for each domain. The independent prompting across domains only requests one single cross-entropy loss for training and one simple K-NN operation as a domain identifier for inference. The learning paradigm derives an image prompt learning approach and a novel language-image prompt learning approach. Owning an excellent scalability (0.03% parameter increase per domain), the best of our approaches achieves a remarkable relative improvement (an average of about 30%) over the best of the state-of-the-art exemplar-free methods for three standard DIL tasks, and even surpasses the best of them relatively by about 6% in average when they use exemplars. Source code is available at \url{https://github.com/iamwangyabin/S-Prompts}.

Subfields: Artificial Intelligence, Computer Vision and Pattern Recognition, Prompting & Reasoning, Meta & Continual Learning

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Artificial Intelligence

1. Han-Jia Ye
2. De-chuan Zhan
3. Da-Wei Zhou
4. Ziwei Liu
5. Hang Su
6. Jun Zhu
7. Qiwen Wang
8. Liyuan Wang

In Computer Vision and Pattern Recognition

1. Han-Jia Ye
2. De-chuan Zhan
3. Jun Zhu
4. Ziwei Liu
5. Luc Van Gool
6. Bernt Schiele
7. Alan L. Yuille
8. Dacheng Tao

In Prompting & Reasoning

1. Liyuan Wang
2. Jun Zhu
3. De-chuan Zhan
4. Hang Su
5. Da-Wei Zhou
6. Ziwei Liu
7. Qiwen Wang
8. Han-Jia Ye

In Meta & Continual Learning

1. Han-Jia Ye
2. De-chuan Zhan
3. Xingxing Zhang
4. Da-Wei Zhou
5. Liyuan Wang
6. Jun Zhu
7. Hang Su
8. Qiwen Wang

Citations in the corpus

• A Comprehensive Survey of Continual Learning: Theory, Method and Application (2023, methodology): “…optimizing a weighted summation of the prompt pool with attention factors (CODA-Prompt [206]), using explicitly task-sharing and task-specific prompts (DualPrompt [207]) or only task-specific prompts (S-Prompts [208]), progressive expansion of task-specific prompts (Progressive Prompts [209]), etc.”
• CODA-Prompt: COntinual Decomposed Attention-Based Prompting for Rehearsal-Free Continual Learning (2022, methodology): “We note that the recent S-Prompts [65] method is similar to these approaches but designed for domain-incremental learning (i.e., learning the same set of classes under covariate distribution task shifts), which is different than the class-incremental continual learning setting of our paper (i.e., le”
• Steering Prototypes with Prompt-tuning for Rehearsal-free Continual Learning (2023, background): “In a similar vein, S-prompts [57] harnesses domain-specific”
• Deep Class-Incremental Learning: A Survey (2023, methodology): “Apart from pre-trained ViT, S-Prompt [75] utilizes the pre-trained language-vision model CLIP [147] for incremental learning, which simultaneously learns language prompts and visual prompts to boost representative embeddings.”
• InfLoRA: Interference-Free Low-Rank Adaptation for Continual Learning (2024, methodology): “Through (2), we can find that each column of ∂ L ∂ W can be represented as input h multiplied by a real value a k ( 1 ≤ k ≤ d O ). of Therefore, in the linear layer, each column the gradient ∂ L ∂ W lies in the span of input.”
• Class-Incremental Learning: A Survey (2023, methodology): “As a result, algorithms should be designed to address catastrophic forgetting in two aspects ( i.e. , classifier aspect and generative model aspect) when using generative replay.”
• A Comprehensive Survey on Test-Time Adaptation Under Distribution Shifts (2023, neutral): “Another related topic is domain-incremental learning [46, 47] which tackles the same type of problem but in diverse contexts.”
• One Size Fits All for Semantic Shifts: Adaptive Prompt Tuning for Continual Learning (2023, methodology): “Again in Figure 4, the prompt for ˆ G t 1 , 1 is proactively adapted to τ 2 , τ 3 and τ t and reserved, even though it is not yet used for CL.”
• Expandable Subspace Ensemble for Pre-Trained Model-Based Class-Incremental Learning (2024, methodology): “Correspondingly, most meth-ods [49, 60–62] learn a prompt pool to adaptively select the instance-specific prompt [31] for model updating.”