【摘要】 海量的数据和计算机强大的计算能力使深度模型在众多单模态和多模态任务上取得优异性能。当前高性能深度模型通常在静态的学习场景中训练，模型只在全部的数据集上进行一次联合训练。然而，在实际应用中，数据是不断产生的，任务以多批次的形式持续地到达，在这种环境下，深度模型面临动态的学习场景，即增量学习场景。由于无法同时访问所有的旧任务数据，增量学习场景中的深度模型在训练时会面临灾难性遗忘问题。如何缓解灾难性遗忘问题是增量学习领域的重要研究目标。本文围绕增量学习领域的研究进展，从增量学习问题定义、评估指标、增量学习范式和增量学习挑战总结增量学习相关背景；从模型参数正则化、样本重放、模型结构化和预训练模型微调4个角度汇总增量学习最新方法；从语义分割、图像生成和文本生成3个单模态领域应用以及视觉—语言、音频—视觉两个多模态领域应用归纳增量学习应用及相关方法。从国内和国外两个角度对比增量学习领域的科研投入和发展情况，并对增量学习领域的未来发展进行展望。本文可为研究人员和从业人员提供增量学习领域的最新进展。更多还原

【Abstract】 The rapid development of information technology has led to the explosive growth of multimedia data, including text, images, and videos. The vast amount of data and the powerful computational capabilities have significantly driven advances in deep learning, allowing models to achieve exceptional performance in single-modal tasks, such as object detection and semantic segmentation, as well as multimodal tasks, such as cross-modal retrieval and question answering. Highperformance deep learning models are typically developed in static learning scenarios in which they can access all training data simultaneously and update the model using the entire dataset. However, in real-world applications, deep models always learn from task streams that dynamically receive new data over time. In these scenarios, deep models must simultaneously retain old task data and maintain performance on old and new tasks through repeated joint training. This approach of unrestricted joint training and continuous dataset expansion to maintain high performance incurs substantial time and financial costs. Furthermore, the prolonged runtime of high-power devices during joint training leads to significant carbon emissions, thereby contributing to environmental pollution. Such costs are even higher in multimodal scenarios due to the vast amounts of multimodal data and the large number of model parameters. This learning strategy, which is misaligned with the sustainability goals of artificial intelligence(AI), is impractical for real-world applications. Therefore, deep models must be capable of adapting to new tasks in dynamic environments, commonly known as incremental learning scenarios. However, developing a high-performance incremental learning model in incremental learning scenarios remains a challenging task, because the model must update its knowledge base each time it receives new data, without access to previously learned data. Due to the unavailability of previously learned data, models face the problem of catastrophic forgetting(CF) in which they tend to forget previously acquired knowledge when learning new tasks. CF degrades model performance, especially in scenarios with high privacy and security requirements, such as personal data in medical image processing. Therefore, incremental learning methods must possess the ability to acquire new knowledge(plasticity) while simultaneously retaining previously learned knowledge(stability). Enhancing stability reduces plasticity, while increasing plasticity causes instability in old tasks. These conflicting demands form the stability–plasticity dilemma, and resolving it is the primary challenge for researchers specializing in the field of incremental learning. This paper categorizes incremental learning methods into four perspectives: regularization-, replay-, and architecture-based methods, as well as methods based on fine-tuning pre-trained model. Regularization-based incremental learning methods mitigate catastrophic forgetting by adding regularization terms that optimize model parameters, adjusting the update of key parameters related to previous tasks while learning new tasks. Such an approach can be divided into two types: parameter and output regularization. The former adds regularization terms to the model’s parameters, while the latter applies them to the model’s outputs. Replaybased incremental learning methods transfer key knowledge from representative old samples to mitigate catastrophic forgetting. Replay-based methods are further divided into generation-and experience-based replay wherein the former generates old representative samples using generative models, while the latter retains actual old representative samples for training.Structure-based methods mitigate CF by maintaining task-specific model parameters. These methods are subdivided into neuron-expansion and parameter-isolation approaches. Neuron-expansion approaches allocate new model parameters for each task by expanding the network, while parameter-isolation approaches freeze key parameters of old tasks to ensure that learning new tasks does not interfere with previous ones. The fine-tuning of pre-trained model methods apply fine-tuning strategies to large models to learn new tasks directly, while leveraging their generalization ability to maintain performance on old tasks. Such methods can be divided into prompt-and representation-based fine-tuning approaches. The former introduces prompts to fine-tune the pre-trained model, thereby balancing old and new knowledge. In comparison, the latter directly leverages the model’s generalization ability to build classifiers, utilizing high-quality feature representations to handle new tasks without significant changes to the model architecture. This paper also provides a mathematical definition of incremental learning scenarios and the objective function for optimizing incremental learning models. In particular, it summarizes six key evaluation metrics used in the incremental learning field: average accuracy(AA), average incremental accuracy(AIA), average forgetting rate(AF), forward transfer rate(FTR), backward transfer rate(BTR), and learning plasticity(LP). Based on the tasks stream pattern in incremental scenarios, the need for task identification, and the number of classification heads, this paper divides incremental learning into three subfields: task-incremental learning(TIL), domain-incremental learning(DIL), and class-incremental learning(CIL), along with detailed explanations and mathematical definitions. This paper summarizes the latest research progress in single-modal fields, including semantic segmentation, image generation, and large language models, as well as multimodal fields like vision-language and vision-audio multimodal incremental learning, based on the application scenarios of incremental learning. Furthermore, this paper surveys the publication status of papers in leading English journals and major conferences in the field to compare the research levels of incremental learning between domestic and international scholars. The comparison analyzes research investments and progress in incremental learning, both domestically and internationally, based on the total number of published papers and the average citation count per paper. Finally, this article anticipates three future directions for the development of incremental learning: large multimodal model incremental learning, incremental learning based on novel deep network architectures, and continual forgetting of knowledge for AI security.更多还原