Abstract:Domain Incremental Learning (DIL) focuses on handling complex domain shifts of a continuous data stream for visual tasks such as image classification and image segmentation. In real life, severe domain gaps in DIL are generated from various sources such as data style shifts, data quality degradation, environment changes, and so on. The well-known catastrophic forgetting issue in DIL becomes even more critical when simultaneously considering multiple sources of domain shifts. In this paper, we propose a unified and effective paradigm named Compositional Prompting (C-Prompt) to mitigate the critical forgetting challenge in DIL for image classification tasks. Unlike a popular type of conventional DIL approaches that need to retain abundant exemplars from the old domains, our exemplar-free C-Prompt leverages a prompt-guided Batch-wise Exponential Moving Average (BEMA) strategy to adaptively consolidate learned knowledge without retaining any exemplars. A set of prompts shared across different domains is designed to estimate the knowledge shifts for automatically balancing knowledge acquisition and forgetting. To enhance the learning ability, our proposed C-Prompt explores a domain-specific pool of learnable prompts for each domain, and all the prompt pools are further exploited in a cross-domain compositional manner to facilitate inference. Since the latest prompting-based DIL methods aim to learn one individual prompt for each domain, they always suffer from critical performance degradation caused by the incorrect prediction of domain index during inference and the limited learning capacity by using a single prompt per domain. Instead, our C-Prompt can not only readily acquire domain-specific knowledge but also exploit domain-shared knowledge. Extensive experiments on various large-scale multi-domain benchmarks have demonstrated the superiority of our proposed C-Prompt compared with state-of-the-art methods. Code is available at https://github.com/zhoujiahuan1991/IJCV2024-C-Prompt .

INCPrompt: Task-Aware incremental Prompting for Rehearsal-Free Class-incremental Learning

Prompting to Prompt for Rehearsal-Free Class Incremental Learning.

KC-Prompt: End-To-End Knowledge-Complementary Prompting for Rehearsal-Free Continual Learning

When Prompt-based Incremental Learning Does Not Meet Strong Pretraining

Hierarchical Prompts for Rehearsal-free Continual Learning

Consistent Prompting for Rehearsal-Free Continual Learning

CAPrompt: Cyclic Prompt Aggregation for Pre-Trained Model Based Class Incremental Learning

CP-Prompt: Composition-Based Cross-modal Prompting for Domain-Incremental Continual Learning

Multi-level Prompt As Extra-Embedding for Incremental Learning

PECTP: Parameter-Efficient Cross-Task Prompts for Incremental Vision Transformer

Vector Quantization Prompting for Continual Learning

Compositional Prompting for Anti-Forgetting in Domain Incremental Learning

Space-time Prompting for Video Class-incremental Learning

RCS-Prompt: Learning Prompt to Rearrange Class Space for Prompt-based Continual Learning

Prompt Customization for Continual Learning

Evolving Parameterized Prompt Memory for Continual Learning

Dynamically Anchored Prompting for Task-Imbalanced Continual Learning

PEARL: Input-Agnostic Prompt Enhancement with Negative Feedback Regulation for Class-Incremental Learning

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