Continual learning aims to allow models to learn new tasks without forgetting what has been learned before. This work introduces Elastic Variational Continual Learning with Weight Consolidation (EVCL), a novel hybrid model that integrates the variational posterior approximation mechanism of Variational Continual Learning (VCL) with the regularization-based parameter-protection strategy of Elastic Weight Consolidation (EWC). By combining the strengths of both methods, EVCL effectively mitigates catastrophic forgetting and enables better capture of dependencies between model parameters and task-specific data. Evaluated on five discriminative tasks, EVCL consistently outperforms existing baselines in both domain-incremental and task-incremental learning scenarios for deep discriminative models.
Dataset distillation aims to condense large datasets into a small number of synthetic examples that can be used as drop-in replacements when training new models. It has applications to interpretability, neural architecture search, privacy, and continual learning. Despite strong successes in supervised domains, such methods have not yet been extended to reinforcement learning, where the lack of a fixed dataset renders most distillation methods unusable. Filling the gap, we formalize behaviour distillation, a setting that aims to discover and then condense the information required for training an expert policy into a synthetic dataset of state-action pairs, without access to expert data. We then introduce Hallucinating Datasets with Evolution Strategies (HaDES), a method for behaviour distillation that can discover datasets of just four state-action pairs which, under supervised learning, train agents to competitive performance levels in continuous control tasks. We show that these datasets generalize out of distribution to training policies with a wide range of architectures and hyperparameters. We also demonstrate application to a downstream task, namely training multi-task agents in a zero-shot fashion. Beyond behaviour distillation, HaDES provides significant improvements in neuroevolution for RL over previous approaches and achieves SoTA results on one standard supervised dataset distillation task. Finally, we show that visualizing the synthetic datasets can provide human-interpretable task insights.
Video language continual learning involves continuously adapting to information from video and text inputs, enhancing a model's ability to handle new tasks while retaining prior knowledge. This field is a relatively under-explored area, and establishing appropriate datasets is crucial for facilitating communication and research in this field. In this study, we present the first dedicated benchmark, ViLCo-Bench, designed to evaluate continual learning models across a range of video-text tasks. The dataset comprises ten-minute-long videos and corresponding language queries collected from publicly available datasets. Additionally, we introduce a novel memory-efficient framework that incorporates self-supervised learning and mimics long-term and short-term memory effects. This framework addresses challenges including memory complexity from long video clips, natural language complexity from open queries, and text-video misalignment. We posit that ViLCo-Bench, with greater complexity compared to existing continual learning benchmarks, would serve as a critical tool for exploring the video-language domain, extending beyond conventional class-incremental tasks, and addressing complex and limited annotation issues. The curated data, evaluations, and our novel method are available at https://github.com/cruiseresearchgroup/ViLCo.
Adapting Automatic Speech Recognition (ASR) models to new domains leads to Catastrophic Forgetting (CF) of previously learned information. This paper addresses CF in the challenging context of Online Continual Learning (OCL), with tasks presented as a continuous data stream with unknown boundaries. We extend OCL for ASR into the unsupervised realm, by leveraging self-training (ST) to facilitate unsupervised adaptation, enabling models to adapt continually without label dependency and without forgetting previous knowledge. Through comparative analysis of various OCL and ST methods across two domain adaptation experiments, we show that UOCL suffers from significantly less forgetting compared to supervised OCL, allowing UOCL methods to approach the performance levels of supervised OCL. Our proposed UOCL extensions further boosts UOCL's efficacy. Our findings represent a significant step towards continually adaptable ASR systems, capable of leveraging unlabeled data across diverse domains.
In this paper, we introduce \textbf{Share}d \textbf{Lo}w \textbf{R}ank \textbf{A}daptation (ShareLoRA), a Large Language Model (LLM) fine-tuning technique that balances parameter efficiency, adaptability, and robustness without compromising performance. By strategically sharing the low-rank weight matrices across different layers, ShareLoRA achieves 44\% to 96\% reduction in trainable parameters compared to standard LoRA, alongside a substantial decrease in memory overhead. This efficiency gain scales with model size, making ShareLoRA particularly advantageous for resource-constrained environments. Importantly, ShareLoRA not only maintains model performance but also exhibits robustness in both classification and generation tasks across diverse models, including RoBERTa, GPT-2, and LLaMA series (1, 2, and 3). It consistently outperforms LoRA in zero-shot, few-shot, and continual fine-tuning scenarios, achieving up to 1.2\% average accuracy improvement, and enhanced generalization across domains. In continual learning settings, ShareLoRA achieves 1.2\% higher accuracy on GSM8K, 0.6\% on HumanEval, and 0.5\% on both MMLU and MMLU-Pro. Our results demonstrate that ShareLoRA supports high-quality fine-tuning while offering strong generalization and continual adaptation across various model scales and diverse tasks.
Continual learning is a challenging problem in machine learning, especially for image classification tasks with imbalanced datasets. It becomes even more challenging when it involves learning new classes incrementally. One method for incremental class learning, addressing dataset imbalance, is rehearsal using previously stored data. In rehearsal-based methods, access to previous data is required for either training the classifier or the generator, but it may not be feasible due to storage, legal, or data access constraints. Although there are many rehearsal-free alternatives for class incremental learning, such as parameter or loss regularization, knowledge distillation, and dynamic architectures, they do not consistently achieve good results, especially on imbalanced data. This paper proposes a new approach called Data-Free Generative Replay (DFGR) for class incremental learning, where the generator is trained without access to real data. In addition, DFGR also addresses dataset imbalance in continual learning of an image classifier. Instead of using training data, DFGR trains a generator using mean and variance statistics of batch-norm and feature maps derived from a pre-trained classification model. The results of our experiments demonstrate that DFGR performs significantly better than other data-free methods and reveal the performance impact of specific parameter settings. DFGR achieves up to 88.5% and 46.6% accuracy on MNIST and FashionMNIST datasets, respectively. Our code is available at https://github.com/2younis/DFGR
Stream Learning (SL) requires models that can quickly adapt to continuously evolving data, posing significant challenges in both computational efficiency and learning accuracy. Effective data selection is critical in SL to ensure a balance between information retention and training efficiency. Traditional rule-based data selection methods struggle to accommodate the dynamic nature of streaming data, highlighting the necessity for innovative solutions that effectively address these challenges. Recent approaches to handling changing data distributions face challenges that limit their effectiveness in fast-paced environments. In response, we propose StreamFP, a novel approach that uniquely employs dynamic, learnable parameters called fingerprints to enhance data selection efficiency and adaptability in stream learning. StreamFP optimizes coreset selection through its unique fingerprint-guided mechanism for efficient training while ensuring robust buffer updates that adaptively respond to data dynamics, setting it apart from existing methods in stream learning. Experimental results demonstrate that StreamFP outperforms state-of-the-art methods by achieving accuracy improvements of 15.99%, 29.65%, and 51.24% compared to baseline models across varying data arrival rates, alongside a training throughput increase of 4.6x.
Existing prompt-tuning methods have demonstrated impressive performances in continual learning (CL), by selecting and updating relevant prompts in the vision-transformer models. On the contrary, this paper aims to learn each task by tuning the prompts in the direction orthogonal to the subspace spanned by previous tasks' features, so as to ensure no interference on tasks that have been learned to overcome catastrophic forgetting in CL. However, different from the orthogonal projection in the traditional CNN architecture, the prompt gradient orthogonal projection in the ViT architecture shows completely different and greater challenges, i.e., 1) the high-order and non-linear self-attention operation; 2) the drift of prompt distribution brought by the LayerNorm in the transformer block. Theoretically, we have finally deduced two consistency conditions to achieve the prompt gradient orthogonal projection, which provide a theoretical guarantee of eliminating interference on previously learned knowledge via the self-attention mechanism in visual prompt tuning. In practice, an effective null-space-based approximation solution has been proposed to implement the prompt gradient orthogonal projection. Extensive experimental results demonstrate the effectiveness of anti-forgetting on four class-incremental benchmarks with diverse pre-trained baseline models, and our approach achieves superior performances to state-of-the-art methods. Our code is available at https://github.com/zugexiaodui/VPTinNSforCL.
Continual Learning (CL) poses a significant challenge in Artificial Intelligence, aiming to mirror the human ability to incrementally acquire knowledge and skills. While extensive research has focused on CL within the context of classification tasks, the advent of increasingly powerful generative models necessitates the exploration of Continual Learning of Generative models (CLoG). This paper advocates for shifting the research focus from classification-based CL to CLoG. We systematically identify the unique challenges presented by CLoG compared to traditional classification-based CL. We adapt three types of existing CL methodologies, replay-based, regularization-based, and parameter-isolation-based methods to generative tasks and introduce comprehensive benchmarks for CLoG that feature great diversity and broad task coverage. Our benchmarks and results yield intriguing insights that can be valuable for developing future CLoG methods. Additionally, we will release a codebase designed to facilitate easy benchmarking and experimentation in CLoG publicly at https://github.com/linhaowei1/CLoG. We believe that shifting the research focus to CLoG will benefit the continual learning community and illuminate the path for next-generation AI-generated content (AIGC) in a lifelong learning paradigm.
In recent years, continual learning (CL) techniques have made significant progress in learning from streaming data while preserving knowledge across sequential tasks, particularly in the realm of euclidean data. To foster fair evaluation and recognize challenges in CL settings, several evaluation frameworks have been proposed, focusing mainly on the single- and multi-label classification task on euclidean data. However, these evaluation frameworks are not trivially applicable when the input data is graph-structured, as they do not consider the topological structure inherent in graphs. Existing continual graph learning (CGL) evaluation frameworks have predominantly focussed on single-label scenarios in the node classification (NC) task. This focus has overlooked the complexities of multi-label scenarios, where nodes may exhibit affiliations with multiple labels, simultaneously participating in multiple tasks. We develop a graph-aware evaluation (\agale) framework that accommodates both single-labeled and multi-labeled nodes, addressing the limitations of previous evaluation frameworks. In particular, we define new incremental settings and devise data partitioning algorithms tailored to CGL datasets. We perform extensive experiments comparing methods from the domains of continual learning, continual graph learning, and dynamic graph learning (DGL). We theoretically analyze \agale and provide new insights about the role of homophily in the performance of compared methods. We release our framework at https://github.com/Tianqi-py/AGALE.
Continual learning of partially similar tasks poses a challenge for artificial neural networks, as task similarity presents both an opportunity for knowledge transfer and a risk of interference and catastrophic forgetting. However, it remains unclear how task similarity in input features and readout patterns influences knowledge transfer and forgetting, as well as how they interact with common algorithms for continual learning. Here, we develop a linear teacher-student model with latent structure and show analytically that high input feature similarity coupled with low readout similarity is catastrophic for both knowledge transfer and retention. Conversely, the opposite scenario is relatively benign. Our analysis further reveals that task-dependent activity gating improves knowledge retention at the expense of transfer, while task-dependent plasticity gating does not affect either retention or transfer performance at the over-parameterized limit. In contrast, weight regularization based on the Fisher information metric significantly improves retention, regardless of task similarity, without compromising transfer performance. Nevertheless, its diagonal approximation and regularization in the Euclidean space are much less robust against task similarity. We demonstrate consistent results in a permuted MNIST task with latent variables. Overall, this work provides insights into when continual learning is difficult and how to mitigate it.
In the present era of deep learning, continual learning research is mainly focused on mitigating forgetting when training a neural network with stochastic gradient descent on a non-stationary stream of data. On the other hand, in the more classical literature of statistical machine learning, many models have sequential Bayesian update rules that yield the same learning outcome as the batch training, i.e., they are completely immune to catastrophic forgetting. However, they are often overly simple to model complex real-world data. In this work, we adopt the meta-learning paradigm to combine the strong representational power of neural networks and simple statistical models' robustness to forgetting. In our novel meta-continual learning framework, continual learning takes place only in statistical models via ideal sequential Bayesian update rules, while neural networks are meta-learned to bridge the raw data and the statistical models. Since the neural networks remain fixed during continual learning, they are protected from catastrophic forgetting. This approach not only achieves significantly improved performance but also exhibits excellent scalability. Since our approach is domain-agnostic and model-agnostic, it can be applied to a wide range of problems and easily integrated with existing model architectures.
Continual learning with deep neural networks presents challenges distinct from both the fixed-dataset and convex continual learning regimes. One such challenge is plasticity loss, wherein a neural network trained in an online fashion displays a degraded ability to fit new tasks. This problem has been extensively studied in both supervised learning and off-policy reinforcement learning (RL), where a number of remedies have been proposed. Still, plasticity loss has received less attention in the on-policy deep RL setting. Here we perform an extensive set of experiments examining plasticity loss and a variety of mitigation methods in on-policy deep RL. We demonstrate that plasticity loss is pervasive under domain shift in this regime, and that a number of methods developed to resolve it in other settings fail, sometimes even performing worse than applying no intervention at all. In contrast, we find that a class of ``regenerative'' methods are able to consistently mitigate plasticity loss in a variety of contexts, including in gridworld tasks and more challenging environments like Montezuma's Revenge and ProcGen.
Continual learning methods are known to suffer from catastrophic forgetting, a phenomenon that is particularly hard to counter for methods that do not store exemplars of previous tasks. Therefore, to reduce potential drift in the feature extractor, existing exemplar-free methods are typically evaluated in settings where the first task is significantly larger than subsequent tasks. Their performance drops drastically in more challenging settings starting with a smaller first task. To address this problem of feature drift estimation for exemplar-free methods, we propose to adversarially perturb the current samples such that their embeddings are close to the old class prototypes in the old model embedding space. We then estimate the drift in the embedding space from the old to the new model using the perturbed images and compensate the prototypes accordingly. We exploit the fact that adversarial samples are transferable from the old to the new feature space in a continual learning setting. The generation of these images is simple and computationally cheap. We demonstrate in our experiments that the proposed approach better tracks the movement of prototypes in embedding space and outperforms existing methods on several standard continual learning benchmarks as well as on fine-grained datasets. Code is available at https://github.com/dipamgoswami/ADC.
Broad, open source availability of large pretrained foundation models on the internet through platforms such as HuggingFace has taken the world of practical deep learning by storm. A classical pipeline for neural network training now typically consists of finetuning these pretrained network on a small target dataset instead of training from scratch. In the case of large models this can be done even on modest hardware using a low rank training technique known as Low-Rank Adaptation (LoRA). While Low Rank training has already been studied in the continual learning setting, existing works often consider storing the learned adapter along with the existing model but rarely attempt to modify the weights of the pretrained model by merging the LoRA with the existing weights after finishing the training of each task. In this article we investigate this setting and study the impact of LoRA rank on the forgetting of the pretraining foundation task and on the plasticity and forgetting of subsequent ones. We observe that this rank has an important impact on forgetting of both the pretraining and downstream tasks. We also observe that vision transformers finetuned in that way exhibit a sort of ``contextual'' forgetting, a behaviour that we do not observe for residual networks and that we believe has not been observed yet in previous continual learning works.
In autonomous driving, even a meticulously trained model can encounter failures when facing unfamiliar scenarios. One of these scenarios can be formulated as an online continual learning (OCL) problem. That is, data come in an online fashion, and models are updated according to these streaming data. Two major OCL challenges are catastrophic forgetting and data imbalance. To address these challenges, in this paper, we propose an Analytic Exemplar-Free Online Continual Learning algorithm (AEF-OCL). The AEF-OCL leverages analytic continual learning principles and employs ridge regression as a classifier for features extracted by a large backbone network. It solves the OCL problem by recursively calculating the analytical solution, ensuring an equalization between the continual learning and its joint-learning counterpart, and works without the need to save any used samples (i.e., exemplar-free). Additionally, we introduce a Pseudo-Features Generator (PFG) module that recursively estimates the mean and the variance of real features for each class. It over-samples offset pseudo-features from the same normal distribution as the real features, thereby addressing the data imbalance issue. Experimental results demonstrate that despite being an exemplar-free strategy, our method outperforms various methods on the autonomous driving SODA10M dataset. Source code is available at https://github.com/ZHUANGHP/Analytic-continual-learning.
Model editing is an emerging field that focuses on updating the knowledge embedded within large language models (LLMs) without extensive retraining. However, current model editing methods significantly compromise the general abilities of LLMs as the number of edits increases, and this trade-off poses a substantial challenge to the continual learning of LLMs. In this paper, we first theoretically analyze that the factor affecting the general abilities in sequential model editing lies in the condition number of the edited matrix. The condition number of a matrix represents its numerical sensitivity, and therefore can be used to indicate the extent to which the original knowledge associations stored in LLMs are perturbed after editing. Subsequently, statistical findings demonstrate that the value of this factor becomes larger as the number of edits increases, thereby exacerbating the deterioration of general abilities. To this end, a framework termed Perturbation Restraint on Upper bouNd for Editing (PRUNE) is proposed, which applies the condition number restraints in sequential editing. These restraints can lower the upper bound on perturbation to edited models, thus preserving the general abilities. Systematically, we conduct experiments employing three popular editing methods on three LLMs across four representative downstream tasks. Evaluation results show that PRUNE can preserve considerable general abilities while maintaining the editing performance effectively in sequential model editing. The code and data are available at https://github.com/mjy1111/PRUNE.
The goal of Continual Learning (CL) task is to continuously learn multiple new tasks sequentially while achieving a balance between the plasticity and stability of new and old knowledge. This paper analyzes that this insufficiency arises from the ineffective handling of outliers, leading to abnormal gradients and unexpected model updates. To address this issue, we enhance the data-aware and parameter-aware robustness of CL, proposing a Robust Continual Learning (RCL) method. From the data perspective, we develop a contrastive loss based on the concepts of uniformity and alignment, forming a feature distribution that is more applicable to outliers. From the parameter perspective, we present a forward strategy for worst-case perturbation and apply robust gradient projection to the parameters. The experimental results on three benchmarks show that the proposed method effectively maintains robustness and achieves new state-of-the-art (SOTA) results. The code is available at: https://github.com/HanxiXiao/RCL
We introduce for continual learning Autodiff Quadratic Consolidation (AQC), which approximates the previous loss function with a quadratic function, and Neural Consolidation (NC), which approximates the previous loss function with a neural network. Although they are not scalable to large neural networks, they can be used with a fixed pre-trained feature extractor. We empirically study these methods in class-incremental learning, for which regularization-based methods produce unsatisfactory results, unless combined with replay. We find that for small datasets, quadratic approximation of the previous loss function leads to poor results, even with full Hessian computation, and NC could significantly improve the predictive performance, while for large datasets, when used with a fixed pre-trained feature extractor, AQC provides superior predictive performance. We also find that using tanh-output features can improve the predictive performance of AQC. In particular, in class-incremental Split MNIST, when a Convolutional Neural Network (CNN) with tanh-output features is pre-trained on EMNIST Letters and used as a fixed pre-trained feature extractor, AQC can achieve predictive performance comparable to joint training.
For the first time, we explore few-shot tuning of vision foundation models for class-incremental learning. Unlike existing few-shot class incremental learning (FSCIL) methods, which train an encoder on a base session to ensure forward compatibility for future continual learning, foundation models are generally trained on large unlabelled data without such considerations. This renders prior methods from traditional FSCIL incompatible for FSCIL with the foundation model. To this end, we propose Consistency-guided Asynchronous Contrastive Tuning (CoACT), a new approach to continually tune foundation models for new classes in few-shot settings. CoACT comprises three components: (i) asynchronous contrastive tuning, which learns new classes by including LoRA modules in the pre-trained encoder, while enforcing consistency between two asynchronous encoders; (ii) controlled fine-tuning, which facilitates effective tuning of a subset of the foundation model; and (iii) consistency-guided incremental tuning, which enforces additional regularization during later sessions to reduce forgetting of the learned classes. We perform an extensive study on 16 diverse datasets and demonstrate the effectiveness of CoACT, outperforming the best baseline method by 2.47% on average and with up to 12.52% on individual datasets. Additionally, CoACT shows reduced forgetting and robustness in low-shot experiments. As an added bonus, CoACT shows up to 13.5% improvement in standard FSCIL over the current SOTA on benchmark evaluations. We make our code publicly available at https://github.com/ShuvenduRoy/CoACT-FSCIL.