In this paper, we focus on a long-term continual learning (CL) task, where a model learns sequentially from a stream of vast tasks over time, acquiring new knowledge while retaining previously learned information in a manner akin to human learning. Unlike traditional CL settings, long-term CL involves handling a significantly larger number of tasks, which exacerbates the issue of catastrophic forgetting. Our work seeks to address two critical questions: 1) How do existing CL methods perform in the context of long-term CL? and 2) How can we mitigate the catastrophic forgetting that arises from prolonged sequential updates? To tackle these challenges, we propose a novel framework inspired by human memory mechanisms for long-term continual learning (Long-CL). Specifically, we introduce a task-core memory management strategy to efficiently index crucial memories and adaptively update them as learning progresses. Additionally, we develop a long-term memory consolidation mechanism that selectively retains hard and discriminative samples, ensuring robust knowledge retention. To facilitate research in this area, we construct and release two multi-modal and textual benchmarks, MMLongCL-Bench and TextLongCL-Bench, providing a valuable resource for evaluating long-term CL approaches. Experimental results show that Long-CL outperforms the previous state-of-the-art by 7.4\% and 6.5\% AP on the two benchmarks, respectively, demonstrating the effectiveness of our approach.
Continual learning for reinforcement learning agents remains a significant challenge, particularly in preserving and leveraging existing information without an external signal to indicate changes in tasks or environments. In this study, we explore the effectiveness of autoencoders in detecting new tasks and matching observed environments to previously encountered ones. Our approach integrates policy optimization with familiarity autoencoders within an end-to-end continual learning system. This system can recognize and learn new tasks or environments while preserving knowledge from earlier experiences and can selectively retrieve relevant knowledge when re-encountering a known environment. Initial results demonstrate successful continual learning without external signals to indicate task changes or reencounters, showing promise for this methodology.
Loss of plasticity in deep neural networks is the gradual reduction in a model's capacity to incrementally learn and has been identified as a key obstacle to learning in non-stationary problem settings. Recent work has shown that deep linear networks tend to be resilient towards loss of plasticity. Motivated by this observation, we propose Adaptive Linearization (AdaLin), a general approach that dynamically adapts each neuron's activation function to mitigate plasticity loss. Unlike prior methods that rely on regularization or periodic resets, AdaLin equips every neuron with a learnable parameter and a gating mechanism that injects linearity into the activation function based on its gradient flow. This adaptive modulation ensures sufficient gradient signal and sustains continual learning without introducing additional hyperparameters or requiring explicit task boundaries. When used with conventional activation functions like ReLU, Tanh, and GeLU, we demonstrate that AdaLin can significantly improve performance on standard benchmarks, including Random Label and Permuted MNIST, Random Label and Shuffled CIFAR-10, and Class-Split CIFAR-100. Furthermore, its efficacy is shown in more complex scenarios, such as class-incremental learning on CIFAR-100 with a ResNet-18 backbone, and in mitigating plasticity loss in off-policy reinforcement learning agents. We perform a systematic set of ablations that show that neuron-level adaptation is crucial for good performance and analyze a number of metrics in the network that might be correlated to loss of plasticity.
Class Incremental Learning (CIL) based on pre-trained models offers a promising direction for open-world continual learning. Existing methods typically rely on correlation-based strategies, where an image's classification feature is used as a query to retrieve the most related key prompts and select the corresponding value prompts for training. However, these approaches face an inherent limitation: fitting the entire feature space of all tasks with only a few trainable prompts is fundamentally challenging. We propose Predictive Prompting (PrePrompt), a novel CIL framework that circumvents correlation-based limitations by leveraging pre-trained models' natural classification ability to predict task-specific prompts. Specifically, PrePrompt decomposes CIL into a two-stage prediction framework: task-specific prompt prediction followed by label prediction. While theoretically appealing, this framework risks bias toward recent classes due to missing historical data for older classifier calibration. PrePrompt then mitigates this by incorporating feature translation, dynamically balancing stability and plasticity. Experiments across multiple benchmarks demonstrate PrePrompt's superiority over state-of-the-art prompt-based CIL methods. Code available at \href{github.com/libo-huang/preprompt}{github.com/libo-huang/preprompt}.
Whole slide image (WSI) classification has emerged as a powerful tool in computational pathology, but remains constrained by domain shifts, e.g., due to different organs, diseases, or institution-specific variations. To address this challenge, we propose an Attention-based Generative Latent Replay Continual Learning framework (AGLR-CL), in a multiple instance learning (MIL) setup for domain incremental WSI classification. Our method employs Gaussian Mixture Models (GMMs) to synthesize WSI representations and patch count distributions, preserving knowledge of past domains without explicitly storing original data. A novel attention-based filtering step focuses on the most salient patch embeddings, ensuring high-quality synthetic samples. This privacy-aware strategy obviates the need for replay buffers and outperforms other buffer-free counterparts while matching the performance of buffer-based solutions. We validate AGLR-CL on clinically relevant biomarker detection and molecular status prediction across multiple public datasets with diverse centers, organs, and patient cohorts. Experimental results confirm its ability to retain prior knowledge and adapt to new domains, offering an effective, privacy-preserving avenue for domain incremental continual learning in WSI classification.
Continual learning (CL) aims to train models that can learn a sequence of tasks without forgetting previously acquired knowledge. A core challenge in CL is balancing stability -- preserving performance on old tasks -- and plasticity -- adapting to new ones. Recently, large pre-trained models have been widely adopted in CL for their ability to support both, offering strong generalization for new tasks and resilience against forgetting. However, their high computational cost at inference time limits their practicality in real-world applications, especially those requiring low latency or energy efficiency. To address this issue, we explore model compression techniques, including pruning and knowledge distillation (KD), and propose two efficient frameworks tailored for class-incremental learning (CIL), a challenging CL setting where task identities are unavailable during inference. The pruning-based framework includes pre- and post-pruning strategies that apply compression at different training stages. The KD-based framework adopts a teacher-student architecture, where a large pre-trained teacher transfers downstream-relevant knowledge to a compact student. Extensive experiments on multiple CIL benchmarks demonstrate that the proposed frameworks achieve a better trade-off between accuracy and inference complexity, consistently outperforming strong baselines. We further analyze the trade-offs between the two frameworks in terms of accuracy and efficiency, offering insights into their use across different scenarios.
In the context of continual learning, acquiring new knowledge while maintaining previous knowledge presents a significant challenge. Existing methods often use experience replay techniques that store a small portion of previous task data for training. In experience replay approaches, data augmentation has emerged as a promising strategy to further improve the model performance by mixing limited previous task data with sufficient current task data. However, we theoretically and empirically analyze that training with mixed samples from random sample pairs may harm the knowledge of previous tasks and cause greater catastrophic forgetting. We then propose GradMix, a robust data augmentation method specifically designed for mitigating catastrophic forgetting in class-incremental learning. GradMix performs gradient-based selective mixup using a class-based criterion that mixes only samples from helpful class pairs and not from detrimental class pairs for reducing catastrophic forgetting. Our experiments on various real datasets show that GradMix outperforms data augmentation baselines in accuracy by minimizing the forgetting of previous knowledge.
In this thesis, we introduce Bayesian filtering as a principled framework for tackling diverse sequential machine learning problems, including online (continual) learning, prequential (one-step-ahead) forecasting, and contextual bandits. To this end, this thesis addresses key challenges in applying Bayesian filtering to these problems: adaptivity to non-stationary environments, robustness to model misspecification and outliers, and scalability to the high-dimensional parameter space of deep neural networks. We develop novel tools within the Bayesian filtering framework to address each of these challenges, including: (i) a modular framework that enables the development adaptive approaches for online learning; (ii) a novel, provably robust filter with similar computational cost to standard filters, that employs Generalised Bayes; and (iii) a set of tools for sequentially updating model parameters using approximate second-order optimisation methods that exploit the overparametrisation of high-dimensional parametric models such as neural networks. Theoretical analysis and empirical results demonstrate the improved performance of our methods in dynamic, high-dimensional, and misspecified models.
Place recognition plays a significant role in SLAM, robot navigation, and autonomous driving applications. Benefiting from deep learning, the performance of LiDAR place recognition (LPR) has been greatly improved. However, many existing learning-based LPR methods suffer from catastrophic forgetting, which severely harms the performance of LPR on previously trained places after training on a new environment. In this paper, we introduce a continual learning framework for LPR via Knowledge Distillation and Fusion (KDF) to alleviate forgetting. Inspired by the ranking process of place recognition retrieval, we present a ranking-aware knowledge distillation loss that encourages the network to preserve the high-level place recognition knowledge. We also introduce a knowledge fusion module to integrate the knowledge of old and new models for LiDAR place recognition. Our extensive experiments demonstrate that KDF can be applied to different networks to overcome catastrophic forgetting, surpassing the state-of-the-art methods in terms of mean Recall@1 and forgetting score.
Recent advancements in text-to-image diffusion models are hindered by high computational demands, limiting accessibility and scalability. This paper introduces KDC-Diff, a novel stable diffusion framework that enhances efficiency while maintaining image quality. KDC-Diff features a streamlined U-Net architecture with nearly half the parameters of the original U-Net (482M), significantly reducing model complexity. We propose a dual-layered distillation strategy to ensure high-fidelity generation, transferring semantic and structural insights from a teacher to a compact student model while minimizing quality degradation. Additionally, replay-based continual learning is integrated to mitigate catastrophic forgetting, allowing the model to retain prior knowledge while adapting to new data. Despite operating under extremely low computational resources, KDC-Diff achieves state-of-the-art performance on the Oxford Flowers and Butterflies & Moths 100 Species datasets, demonstrating competitive metrics such as FID, CLIP, and LPIPS. Moreover, it significantly reduces inference time compared to existing models. These results establish KDC-Diff as a highly efficient and adaptable solution for text-to-image generation, particularly in computationally constrained environments.
This study aims to address the problem of multi-domain task incremental learning~(MTIL), which requires that vision-language models~(VLMs) continuously acquire new knowledge while maintaining their inherent zero-shot recognition capability. Existing paradigms delegate the testing of unseen-domain samples to the original CLIP, which only prevents the degradation of the model's zero-shot capability but fails to enhance the generalization of the VLM further. To this end, we propose a novel MTIL framework, named AFA, which comprises two core modules: (1) an against forward-forgetting adapter that learns task-invariant information for each dataset in the incremental tasks to enhance the zero-shot recognition ability of VLMs; (2) an against backward-forgetting adapter that strengthens the few-shot learning capability of VLMs while supporting incremental learning. Extensive experiments demonstrate that the AFA method significantly outperforms existing state-of-the-art approaches, especially in few-shot MTIL tasks, and surpasses the inherent zero-shot performance of CLIP in terms of transferability. The code is provided in the Supplementary Material.
Continual learning (CL) aims to learn a sequence of tasks without forgetting prior knowledge, but gradient updates for a new task often overwrite the weights learned earlier, causing catastrophic forgetting (CF). We propose Prototype-Augmented Hypernetworks (PAH), a framework where a single hypernetwork, conditioned on learnable task prototypes, dynamically generates task-specific classifier heads on demand. To mitigate forgetting, PAH combines cross-entropy with dual distillation losses, one to align logits and another to align prototypes, ensuring stable feature representations across tasks. Evaluations on Split-CIFAR100 and TinyImageNet demonstrate that PAH achieves state-of-the-art performance, reaching 74.5 % and 63.7 % accuracy with only 1.7 % and 4.4 % forgetting, respectively, surpassing prior methods without storing samples or heads.
Despite advances in spiking neural networks (SNNs) in numerous tasks, their architectures remain highly similar to traditional artificial neural networks (ANNs), restricting their ability to mimic natural connections between biological neurons. This paper develops a new modeling paradigm for SNN with random graph architecture (RGA), termed Cognition-aware SNN (CogniSNN). Furthermore, we improve the expandability and neuroplasticity of CogniSNN by introducing a modified spiking residual neural node (ResNode) to counteract network degradation in deeper graph pathways, as well as a critical path-based algorithm that enables CogniSNN to perform continual learning on new tasks leveraging the features of the data and the RGA learned in the old task. Experiments show that CogniSNN with re-designed ResNode performs outstandingly in neuromorphic datasets with fewer parameters, achieving 95.5% precision in the DVS-Gesture dataset with only 5 timesteps. The critical path-based approach decreases 3% to 5% forgetting while maintaining expected performance in learning new tasks that are similar to or distinct from the old ones. This study showcases the potential of RGA-based SNN and paves a new path for biologically inspired networks based on graph theory.
This article presents a concept-centric paradigm for building agents that can learn continually and reason flexibly. The concept-centric agent utilizes a vocabulary of neuro-symbolic concepts. These concepts, such as object, relation, and action concepts, are grounded on sensory inputs and actuation outputs. They are also compositional, allowing for the creation of novel concepts through their structural combination. To facilitate learning and reasoning, the concepts are typed and represented using a combination of symbolic programs and neural network representations. Leveraging such neuro-symbolic concepts, the agent can efficiently learn and recombine them to solve various tasks across different domains, ranging from 2D images, videos, 3D scenes, and robotic manipulation tasks. This concept-centric framework offers several advantages, including data efficiency, compositional generalization, continual learning, and zero-shot transfer.
This technical report describes an experiment on autoregressive pre-training of Gemma2 2 billion parameter large language model (LLM) with 10\% on the Lithuanian language component of CulturaX from the point of view of continual learning. We apply elastic weight consolidation (EWC) to the full set of the model's parameters and investigate language understanding benchmarks, consisting of Arc, Belebele, Gsm8K, Hellaswag, MMLU, TruthfulQA, and Winogrande sets (both in English and Lithuanian versions), and perplexity benchmarks. We empirically demonstrate that EWC regularisation allows us not only to mitigate catastrophic forgetting effects but also that it is potentially beneficial for learning of the new task with LLMs.
Continual Learning entails progressively acquiring knowledge from new data while retaining previously acquired knowledge, thereby mitigating ``Catastrophic Forgetting'' in neural networks. Our work presents a novel uncertainty-driven Unsupervised Continual Learning framework using Generative Replay, namely ``Replay to Remember (R2R)''. The proposed R2R architecture efficiently uses unlabelled and synthetic labelled data in a balanced proportion using a cluster-level uncertainty-driven feedback mechanism and a VLM-powered generative replay module. Unlike traditional memory-buffer methods that depend on pretrained models and pseudo-labels, our R2R framework operates without any prior training. It leverages visual features from unlabeled data and adapts continuously using clustering-based uncertainty estimation coupled with dynamic thresholding. Concurrently, a generative replay mechanism along with DeepSeek-R1 powered CLIP VLM produces labelled synthetic data representative of past experiences, resembling biological visual thinking that replays memory to remember and act in new, unseen tasks. Extensive experimental analyses are carried out in CIFAR-10, CIFAR-100, CINIC-10, SVHN and TinyImageNet datasets. Our proposed R2R approach improves knowledge retention, achieving a state-of-the-art performance of 98.13%, 73.06%, 93.41%, 95.18%, 59.74%, respectively, surpassing state-of-the-art performance by over 4.36%.
The proliferation of end devices has led to a distributed computing paradigm, wherein on-device machine learning models continuously process diverse data generated by these devices. The dynamic nature of this data, characterized by continuous changes or data drift, poses significant challenges for on-device models. To address this issue, continual learning (CL) is proposed, enabling machine learning models to incrementally update their knowledge and mitigate catastrophic forgetting. However, the traditional centralized approach to CL is unsuitable for end devices due to privacy and data volume concerns. In this context, federated continual learning (FCL) emerges as a promising solution, preserving user data locally while enhancing models through collaborative updates. Aiming at the challenges of limited storage resources for CL, poor autonomy in task shift detection, and difficulty in coping with new adversarial tasks in FCL scenario, we propose a novel FCL framework named SacFL. SacFL employs an Encoder-Decoder architecture to separate task-robust and task-sensitive components, significantly reducing storage demands by retaining lightweight task-sensitive components for resource-constrained end devices. Moreover, $\rm{SacFL}$ leverages contrastive learning to introduce an autonomous data shift detection mechanism, enabling it to discern whether a new task has emerged and whether it is a benign task. This capability ultimately allows the device to autonomously trigger CL or attack defense strategy without additional information, which is more practical for end devices. Comprehensive experiments conducted on multiple text and image datasets, such as Cifar100 and THUCNews, have validated the effectiveness of $\rm{SacFL}$ in both class-incremental and domain-incremental scenarios. Furthermore, a demo system has been developed to verify its practicality.
Class-Incremental Learning (CIL) is a critical capability for real-world applications, enabling learning systems to adapt to new tasks while retaining knowledge from previous ones. Recent advancements in pre-trained models (PTMs) have significantly advanced the field of CIL, demonstrating superior performance over traditional methods. However, understanding how features evolve and are distributed across incremental tasks remains an open challenge. In this paper, we propose a novel approach to modeling feature evolution in PTM-based CIL through the lens of neural collapse (NC), a striking phenomenon observed in the final phase of training, which leads to a well-separated, equiangular feature space. We explore the connection between NC and CIL effectiveness, showing that aligning feature distributions with the NC geometry enhances the ability to capture the dynamic behavior of continual learning. Based on this insight, we introduce Neural Collapse-inspired Pre-Trained Model-based CIL (NCPTM-CIL), a method that dynamically adjusts the feature space to conform to the elegant NC structure, thereby enhancing the continual learning process. Extensive experiments demonstrate that NCPTM-CIL outperforms state-of-the-art methods across four benchmark datasets. Notably, when initialized with ViT-B/16-IN1K, NCPTM-CIL surpasses the runner-up method by 6.73% on VTAB, 1.25% on CIFAR-100, and 2.5% on OmniBenchmark.
Bayesian Federated Learning (BFL) enables uncertainty quantification and robust adaptation in distributed learning. In contrast to the frequentist approach, it estimates the posterior distribution of a global model, offering insights into model reliability. However, current BFL methods neglect continual learning challenges in dynamic environments where data distributions shift over time. We propose a continual BFL framework applied to human sensing with radar data collected over several days. Using Stochastic Gradient Langevin Dynamics (SGLD), our approach sequentially updates the model, leveraging past posteriors to construct the prior for the new tasks. We assess the accuracy, the expected calibration error (ECE) and the convergence speed of our approach against several baselines. Results highlight the effectiveness of continual Bayesian updates in preserving knowledge and adapting to evolving data.
Many applications of computer vision require the ability to adapt to novel data distributions after deployment. Adaptation requires algorithms capable of continual learning (CL). Continual learners must be plastic to adapt to novel tasks while minimizing forgetting of previous tasks.However, CL opens up avenues for noise to enter the training pipeline and disrupt the CL. This work focuses on label noise and instance noise in the context of class-incremental learning (CIL), where new classes are added to a classifier over time, and there is no access to external data from past classes. We aim to understand the sensitivity of CL methods that work by replaying items from a memory constructed using the idea of Coresets. We derive a new bound for the robustness of such a method to uncorrelated instance noise under a general additive noise threat model, revealing several insights. Putting the theory into practice, we create two continual learning algorithms to construct noise-tolerant replay buffers. We empirically compare the effectiveness of prior memory-based continual learners and the proposed algorithms under label and uncorrelated instance noise on five diverse datasets. We show that existing memory-based CL are not robust whereas the proposed methods exhibit significant improvements in maximizing classification accuracy and minimizing forgetting in the noisy CIL setting.