Continual learning aims to enable models to adapt to new datasets without losing performance on previously learned data, often assuming that prior data is no longer available. However, in many practical scenarios, both old and new data are accessible. In such cases, good performance on both datasets is typically achieved by abandoning the model trained on the previous data and re-training a new model from scratch on both datasets. This training from scratch is computationally expensive. In contrast, methods that leverage the previously trained model and old data are worthy of investigation, as they could significantly reduce computational costs. Our evaluation framework quantifies the computational savings of such methods while maintaining or exceeding the performance of training from scratch. We identify key optimization aspects -- initialization, regularization, data selection, and hyper-parameters -- that can each contribute to reducing computational costs. For each aspect, we propose effective first-step methods that already yield substantial computational savings. By combining these methods, we achieve up to 2.7x reductions in computation time across various computer vision tasks, highlighting the potential for further advancements in this area.

Summary and Contributions: The paper presents an algorithm for k-clustering in a sliding window streaming model, where k-clustering means the generalization of k-median and k-means to any fixed l_p-norm. The main theoretical result is an algorithm that achieves O(1)-approximation for points in arbitrary metric space and thus includes the prevalent of Euclidean metric, which is also used in the experimental evaluation. This algorithm is for sliding window streaming, where the algorithm repeatedly solves the clustering problem on the w most recent points in the stream (for parameter w). While the minimal requirement is to estimate the cost of a k-clustering, this algorithm also reports k center points. The usual motivation for this model is to allow old data to expire, and analyze only recent data.

Online continual learning (OCL) involves deep neural networks retaining knowledge from old data while adapting to new data, which is accessible only once. A critical challenge in OCL is catastrophic forgetting, reflected in reduced model performance on old data. Existing replay-based methods mitigate forgetting by replaying buffered samples from old data and learning current samples of new data. In this work, we dissect existing methods and empirically discover that learning and replaying in the same feature space is not conducive to addressing the forgetting issue. Since the learned features associated with old data are readily changed by the features related to new data due to data imbalance, leading to the forgetting problem. Based on this observation, we intuitively explore learning and replaying in different feature spaces. Learning in a feature subspace is sufficient to capture novel knowledge from new data while replaying in a larger feature space provides more feature space to maintain historical knowledge from old data. To this end, we propose a novel OCL approach called experience replay with feature subspace learning (ER-FSL). Firstly, ER-FSL divides the entire feature space into multiple subspaces, with each subspace used to learn current samples. Moreover, it introduces a subspace reuse mechanism to address situations where no blank subspaces exist. Secondly, ER-FSL replays previous samples using an accumulated space comprising all learned subspaces. Extensive experiments on three datasets demonstrate the superiority of ER-FSL over various state-of-the-art methods.

Instance-incremental learning (IIL) focuses on learning continually with data of the same classes. Compared to class-incremental learning (CIL), the IIL is seldom explored because IIL suffers less from catastrophic forgetting (CF). However, besides retaining knowledge, in real-world deployment scenarios where the class space is always predefined, continual and cost-effective model promotion with the potential unavailability of previous data is a more essential demand. Therefore, we first define a new and more practical IIL setting as promoting the model's performance besides resisting CF with only new observations. Two issues have to be tackled in the new IIL setting: 1) the notorious catastrophic forgetting because of no access to old data, and 2) broadening the existing decision boundary to new observations because of concept drift. To tackle these problems, our key insight is to moderately broaden the decision boundary to fail cases while retain old boundary. Hence, we propose a novel decision boundary-aware distillation method with consolidating knowledge to teacher to ease the student learning new knowledge. We also establish the benchmarks on existing datasets Cifar-100 and ImageNet. Notably, extensive experiments demonstrate that the teacher model can be a better incremental learner than the student model, which overturns previous knowledge distillation-based methods treating student as the main role.

Recommender models play a vital role in various industrial scenarios, while often faced with the catastrophic forgetting problem caused by the fast shifting data distribution, e.g., the evolving user interests, click signals fluctuation during sales promotions, etc. To alleviate this problem, a common approach is to reuse knowledge from the historical data. However, preserving the vast and fast-accumulating data is hard, which causes dramatic storage overhead. Memorizing old data through a parametric knowledge base is then proposed, which compresses the vast amount of raw data into model parameters. Despite the flexibility, how to improve the memorization and generalization capabilities of the parametric knowledge base is challenging. In this paper, we propose two constraints to extract Essential and Disentangled Knowledge from past data for rational and generalized recommendation enhancement, which improves the capabilities of the parametric knowledge base without increasing the size of it. The essential principle helps to compress the input into representative vectors that capture the task-relevant information and filter out the noisy information. The disentanglement principle reduces the redundancy of stored information and pushes the knowledge base to focus on capturing the disentangled invariant patterns. These two rules together promote rational compression of information for robust and generalized knowledge representations. Extensive experiments on two datasets justify the effectiveness of the proposed method.

Class-Incremental Learning (CIL) is a practical and challenging problem for achieving general artificial intelligence. Recently, Pre-Trained Models (PTMs) have led to breakthroughs in both visual and natural language processing tasks. Despite recent studies showing PTMs' potential ability to learn sequentially, a plethora of work indicates the necessity of alleviating the catastrophic forgetting of PTMs. Through a pilot study and a causal analysis of CIL, we reveal that the crux lies in the imbalanced causal effects between new and old data. Specifically, the new data encourage models to adapt to new classes while hindering the adaptation of old classes. Similarly, the old data encourages models to adapt to old classes while hindering the adaptation of new classes. In other words, the adaptation process between new and old classes conflicts from the causal perspective. To alleviate this problem, we propose Balancing the Causal Effects (BaCE) in CIL. Concretely, BaCE proposes two objectives for building causal paths from both new and old data to the prediction of new and classes, respectively. In this way, the model is encouraged to adapt to all classes with causal effects from both new and old data and thus alleviates the causal imbalance problem. We conduct extensive experiments on continual image classification, continual text classification, and continual named entity recognition. Empirical results show that BaCE outperforms a series of CIL methods on different tasks and settings.

The researchers taught the robot, called Mobile ALOHA (an acronym for "a low-cost open-source hardware teleoperation system for bimanual operation"), seven different tasks requiring a variety of mobility and dexterity skills, such as rinsing a pan or giving someone a high five. To teach the robot how to cook shrimp, for example, the researchers remotely operated it 20 times to get the shrimp into the plan, flip it, and then serve it. They did it slightly differently each time so the robot learned different ways to do the same task, says Zipeng Fu, a PhD Student at Stanford, who was project co-lead. The robot was then trained on these demonstrations, as well as other human-operated demonstrations for different types of tasks that have nothing to do with shrimp cooking, such as tearing off a paper towel or tape collected by an earlier ALOHA robot without wheels, says Chelsea Finn, an assistant professor at Stanford University, who was an advisor for the project. This "co-training" approach, in which new and old data are combined, helped Mobile ALOHA learn new jobs relatively quickly, compared with the usual approach of training AI systems on thousands if not millions of examples.

The typical model development lifecycle consists of four phases: 1) problem scoping, 2) data definition and collection, 3) model training and iterative improvement through error analysis, and 4) model deployment in production and implementation of continuous monitoring and retraining [1]. While the first three phases are typically performed in an offline setting, model deployment represents the critical step where the ML model becomes available in a production environment, a live application, where it needs to process live data and ideally sustain performance over time to keep delivering value. Model monitoring refers to the process of evaluating the quality of the production data and the performance of the model according to relevant metrics over time. When either data quality or model performance does not meet predefined criteria, a monitoring warning can be triggered, to alert the model owners. Defining an effective model monitoring and retraining strategy is key to successful ML model deployment since it can safeguard model quality over prolonged periods of time.

Contrastive Learning (CL) has been proved to be a powerful self-supervised approach for a wide range of domains, including computer vision and graph representation learning. However, the incremental learning issue of CL has rarely been studied, which brings the limitation in applying it to real-world applications. Contrastive learning identifies the samples with the negative ones from the noise distribution that changes in the incremental scenarios. Therefore, only fitting the change of data without noise distribution causes bias, and directly retraining results in low efficiency. To bridge this research gap, we propose a self-supervised Incremental Contrastive Learning (ICL) framework consisting of (i) a novel Incremental InfoNCE (NCE-II) loss function by estimating the change of noise distribution for old data to guarantee no bias with respect to the retraining, (ii) a meta-optimization with deep reinforced Learning Rate Learning (LRL) mechanism which can adaptively learn the learning rate according to the status of the training processes and achieve fast convergence which is critical for incremental learning. Theoretically, the proposed ICL is equivalent to retraining, which is based on solid mathematical derivation. In practice, extensive experiments in different domains demonstrate that, without retraining a new model, ICL achieves up to 16.7x training speedup and 16.8x faster convergence with competitive results.

Continual Learning for Named Entity Recognition (CL-NER) aims to learn a growing number of entity types over time from a stream of data. However, simply learning Other-Class in the same way as new entity types amplifies the catastrophic forgetting and leads to a substantial performance drop. The main cause behind this is that Other-Class samples usually contain old entity types, and the old knowledge in these Other-Class samples is not preserved properly. Thanks to the causal inference, we identify that the forgetting is caused by the missing causal effect from the old data. To this end, we propose a unified causal framework to retrieve the causality from both new entity types and Other-Class. Furthermore, we apply curriculum learning to mitigate the impact of label noise and introduce a self-adaptive weight for balancing the causal effects between new entity types and Other-Class. Experimental results on three benchmark datasets show that our method outperforms the state-of-the-art method by a large margin. Moreover, our method can be combined with the existing state-of-the-art methods to improve the performance in CL-NER