Despite the great success of pre-trained language models, it is still a challenge to use these models for continual learning, especially for the class-incremental learning (CIL) setting due to catastrophic forgetting (CF). This paper reports our finding that if we formulate CIL as a continual label generation problem, CF is drastically reduced and the generalizable representations of pre-trained models can be better retained. We thus propose a new CIL method (VAG) that also leverages the sparsity of vocabulary to focus the generation and creates pseudo-replay samples by using label semantics. Experimental results show that VAG outperforms baselines by a large margin.
Automatic speech recognition (ASR) models with low-footprint are increasingly being deployed on edge devices for conversational agents, which enhances privacy. We study the problem of federated continual incremental learning for recurrent neural network-transducer (RNN-T) ASR models in the privacy-enhancing scheme of learning on-device, without access to ground truth human transcripts or machine transcriptions from a stronger ASR model. In particular, we study the performance of a self-learning based scheme, with a paired teacher model updated through an exponential moving average of ASR models. Further, we propose using possibly noisy weak-supervision signals such as feedback scores and natural language understanding semantics determined from user behavior across multiple turns in a session of interactions with the conversational agent. These signals are leveraged in a multi-task policy-gradient training approach to improve the performance of self-learning for ASR. Finally, we show how catastrophic forgetting can be mitigated by combining on-device learning with a memory-replay approach using selected historical datasets. These innovations allow for 10% relative improvement in WER on new use cases with minimal degradation on other test sets in the absence of strong-supervision signals such as ground-truth transcriptions.
In the scenario of class-incremental learning (CIL), deep neural networks have to adapt their model parameters to non-stationary data distributions, e.g., the emergence of new classes over time. However, CIL models are challenged by the well-known catastrophic forgetting phenomenon. Typical methods such as rehearsal-based ones rely on storing exemplars of old classes to mitigate catastrophic forgetting, which limits real-world applications considering memory resources and privacy issues. In this paper, we propose a novel rehearsal-free CIL approach that learns continually via the synergy between two Complementary Learning Subnetworks. Our approach involves jointly optimizing a plastic CNN feature extractor and an analytical feed-forward classifier. The inaccessibility of historical data is tackled by holistically controlling the parameters of a well-trained model, ensuring that the decision boundary learned fits new classes while retaining recognition of previously learned classes. Specifically, the trainable CNN feature extractor provides task-dependent knowledge separately without interference; and the final classifier integrates task-specific knowledge incrementally for decision-making without forgetting. In each CIL session, it accommodates new tasks by attaching a tiny set of declarative parameters to its backbone, in which only one matrix per task or one vector per class is kept for knowledge retention. Extensive experiments on a variety of task sequences show that our method achieves competitive results against state-of-the-art methods, especially in accuracy gain, memory cost, training efficiency, and task-order robustness. Furthermore, to make the non-growing backbone (i.e., a model with limited network capacity) suffice to train on more incoming tasks, a graceful forgetting implementation on previously learned trivial tasks is empirically investigated.
Machine Learning (ML) models struggle with data that changes over time or across domains due to factors such as noise, occlusion, illumination, or frequency, unlike humans who can learn from such non independent and identically distributed data. Consequently, a Continual Learning (CL) approach is indispensable, particularly, Domain-Incremental Learning. In this paper, we propose a novel pipeline for identifying tasks in domain-incremental learning scenarios without supervision. The pipeline comprises four steps. First, we obtain base embeddings from the raw data using an existing transformer-based model. Second, we group the embedding densities based on their similarity to obtain the nearest points to each cluster centroid. Third, we train an incremental task classifier using only these few points. Finally, we leverage the lightweight computational requirements of the pipeline to devise an algorithm that decides in an online fashion when to learn a new task using the task classifier and a drift detector. We conduct experiments using the SODA10M real-world driving dataset and several CL strategies. We demonstrate that the performance of these CL strategies with our pipeline can match the ground-truth approach, both in classical experiments assuming task boundaries, and also in more realistic task-agnostic scenarios that require detecting new tasks on-the-fly
Word frequency is a strong predictor in most lexical processing tasks. Thus, any model of word recognition needs to account for how word frequency effects arise. The Discriminative Lexicon Model (DLM; Baayen et al., 2018a, 2019) models lexical processing with linear mappings between words' forms and their meanings. So far, the mappings can either be obtained incrementally via error-driven learning, a computationally expensive process able to capture frequency effects, or in an efficient, but frequency-agnostic solution modelling the theoretical endstate of learning (EL) where all words are learned optimally. In this study we show how an efficient, yet frequency-informed mapping between form and meaning can be obtained (Frequency-informed learning; FIL). We find that FIL well approximates an incremental solution while being computationally much cheaper. FIL shows a relatively low type- and high token-accuracy, demonstrating that the model is able to process most word tokens encountered by speakers in daily life correctly. We use FIL to model reaction times in the Dutch Lexicon Project (Keuleers et al., 2010) and find that FIL predicts well the S-shaped relationship between frequency and the mean of reaction times but underestimates the variance of reaction times for low frequency words. FIL is also better able to account for priming effects in an auditory lexical decision task in Mandarin Chinese (Lee, 2007), compared to EL. Finally, we used ordered data from CHILDES (Brown, 1973; Demuth et al., 2006) to compare mappings obtained with FIL and incremental learning. The mappings are highly correlated, but with FIL some nuances based on word ordering effects are lost. Our results show how frequency effects in a learning model can be simulated efficiently, and raise questions about how to best account for low-frequency words in cognitive models.
Few-shot class-incremental learning (FSCIL) aims to continually learn new classes using a few samples while not forgetting the old classes. The key of this task is effective knowledge transfer from the base session to the incremental sessions. Despite the advance of existing FSCIL methods, the proposed knowledge transfer learning schemes are sub-optimal due to the insufficient optimization for the model's plasticity. To address this issue, we propose a Random Episode Sampling and Augmentation (RESA) strategy that relies on diverse pseudo incremental tasks as agents to achieve the knowledge transfer. Concretely, RESA mimics the real incremental setting and constructs pseudo incremental tasks globally and locally, where the global pseudo incremental tasks are designed to coincide with the learning objective of FSCIL and the local pseudo incremental tasks are designed to improve the model's plasticity, respectively. Furthermore, to make convincing incremental predictions, we introduce a complementary model with a squared Euclidean-distance classifier as the auxiliary module, which couples with the widely used cosine classifier to form our whole architecture. By such a way, equipped with model decoupling strategy, we can maintain the model's stability while enhancing the model's plasticity. Extensive quantitative and qualitative experiments on three popular FSCIL benchmark datasets demonstrate that our proposed method, named Knowledge Transfer-driven Relation Complementation Network (KT-RCNet), outperforms almost all prior methods. More precisely, the average accuracy of our proposed KT-RCNet outperforms the second-best method by a margin of 5.26%, 3.49%, and 2.25% on miniImageNet, CIFAR100, and CUB200, respectively. Our code is available at https://github.com/YeZiLaiXi/KT-RCNet.git.
Multi-view learning (MVL) has gained great success in integrating information from multiple perspectives of a dataset to improve downstream task performance. To make MVL methods more practical in an open-ended environment, this paper investigates a novel paradigm called multi-view class incremental learning (MVCIL), where a single model incrementally classifies new classes from a continual stream of views, requiring no access to earlier views of data. However, MVCIL is challenged by the catastrophic forgetting of old information and the interference with learning new concepts. To address this, we first develop a randomization-based representation learning technique serving for feature extraction to guarantee their separate view-optimal working states, during which multiple views belonging to a class are presented sequentially; Then, we integrate them one by one in the orthogonality fusion subspace spanned by the extracted features; Finally, we introduce selective weight consolidation for learning-without-forgetting decision-making while encountering new classes. Extensive experiments on synthetic and real-world datasets validate the effectiveness of our approach.
Stock trend forecasting is a fundamental task of quantitative investment where precise predictions of price trends are indispensable. As an online service, stock data continuously arrive over time. It is practical and efficient to incrementally update the forecast model with the latest data which may reveal some new patterns recurring in the future stock market. However, incremental learning for stock trend forecasting still remains under-explored due to the challenge of distribution shifts (a.k.a. concept drifts). With the stock market dynamically evolving, the distribution of future data can slightly or significantly differ from incremental data, hindering the effectiveness of incremental updates. To address this challenge, we propose DoubleAdapt, an end-to-end framework with two adapters, which can effectively adapt the data and the model to mitigate the effects of distribution shifts. Our key insight is to automatically learn how to adapt stock data into a locally stationary distribution in favor of profitable updates. Complemented by data adaptation, we can confidently adapt the model parameters under mitigated distribution shifts. We cast each incremental learning task as a meta-learning task and automatically optimize the adapters for desirable data adaptation and parameter initialization. Experiments on real-world stock datasets demonstrate that DoubleAdapt achieves state-of-the-art predictive performance and shows considerable efficiency.
We identify incremental learning dynamics in transformers, where the difference between trained and initial weights progressively increases in rank. We rigorously prove this occurs under the simplifying assumptions of diagonal weight matrices and small initialization. Our experiments support the theory and also show that phenomenon can occur in practice without the simplifying assumptions.
In collaborative learning with streaming data, nodes (e.g., organizations) jointly and continuously learn a machine learning (ML) model by sharing the latest model updates computed from their latest streaming data. For the more resourceful nodes to be willing to share their model updates, they need to be fairly incentivized. This paper explores an incentive design that guarantees fairness so that nodes receive rewards commensurate to their contributions. Our approach leverages an explore-then-exploit formulation to estimate the nodes' contributions (i.e., exploration) for realizing our theoretically guaranteed fair incentives (i.e., exploitation). However, we observe a "rich get richer" phenomenon arising from the existing approaches to guarantee fairness and it discourages the participation of the less resourceful nodes. To remedy this, we additionally preserve asymptotic equality, i.e., less resourceful nodes achieve equal performance eventually to the more resourceful/"rich" nodes. We empirically demonstrate in two settings with real-world streaming data: federated online incremental learning and federated reinforcement learning, that our proposed approach outperforms existing baselines in fairness and learning performance while remaining competitive in preserving equality.
Pre-trained deep neural networks (DNNs) are being widely deployed by industry for making business decisions and to serve users; however, a major problem is model decay, where the DNN's predictions become more erroneous over time, resulting in revenue loss or unhappy users. To mitigate model decay, DNNs are retrained from scratch using old and new data. This is computationally expensive, so retraining happens only once performance significantly decreases. Here, we study how continual learning (CL) could potentially overcome model decay in large pre-trained DNNs and greatly reduce computational costs for keeping DNNs up-to-date. We identify the "stability gap" as a major obstacle in our setting. The stability gap refers to a phenomenon where learning new data causes large drops in performance for past tasks before CL mitigation methods eventually compensate for this drop. We test two hypotheses to investigate the factors influencing the stability gap and identify a method that vastly reduces this gap. In large-scale experiments for both easy and hard CL distributions (e.g., class incremental learning), we demonstrate that our method reduces the stability gap and greatly increases computational efficiency. Our work aligns CL with the goals of the production setting, where CL is needed for many applications.
A major focus of clinical imaging workflow is disease diagnosis and management, leading to medical imaging datasets strongly tied to specific clinical objectives. This scenario has led to the prevailing practice of developing task-specific segmentation models, without gaining insights from widespread imaging cohorts. Inspired by the training program of medical radiology residents, we propose a shift towards universal medical image segmentation, a paradigm aiming to build medical image understanding foundation models by leveraging the diversity and commonality across clinical targets, body regions, and imaging modalities. Towards this goal, we develop Hermes, a novel context-prior learning approach to address the challenges of data heterogeneity and annotation differences in medical image segmentation. In a large collection of eleven diverse datasets (2,438 3D images) across five modalities (CT, PET, T1, T2 and cine MRI) and multiple body regions, we demonstrate the merit of the universal paradigm over the traditional paradigm on addressing multiple tasks within a single model. By exploiting the synergy across tasks, Hermes achieves state-of-the-art performance on all testing datasets and shows superior model scalability. Results on two additional datasets reveals Hermes' strong performance for transfer learning, incremental learning, and generalization to downstream tasks. Hermes's learned priors demonstrate an appealing trait to reflect the intricate relations among tasks and modalities, which aligns with the established anatomical and imaging principles in radiology. The code is available: https://github.com/yhygao/universal-medical-image-segmentation.
Rehearsal-based video incremental learning often employs knowledge distillation to mitigate catastrophic forgetting of previously learned data. However, this method faces two major challenges for video task: substantial computing resources from loading teacher model and limited replay capability from performance-limited teacher model. To address these problems, we first propose a knowledge distillation-free framework for rehearsal-based video incremental learning called \textit{Teacher Agent}. Instead of loading parameter-heavy teacher networks, we introduce an agent generator that is either parameter-free or uses only a few parameters to obtain accurate and reliable soft labels. This method not only greatly reduces the computing requirement but also circumvents the problem of knowledge misleading caused by inaccurate predictions of the teacher model. Moreover, we put forward a self-correction loss which provides an effective regularization signal for the review of old knowledge, which in turn alleviates the problem of catastrophic forgetting. Further, to ensure that the samples in the memory buffer are memory-efficient and representative, we introduce a unified sampler for rehearsal-based video incremental learning to mine fixed-length key video frames. Interestingly, based on the proposed strategies, the network exhibits a high level of robustness against spatial resolution reduction when compared to the baseline. Extensive experiments demonstrate the advantages of our method, yielding significant performance improvements while utilizing only half the spatial resolution of video clips as network inputs in the incremental phases.
Deep learning (DL) models for segmenting various anatomical structures have achieved great success via a static DL model that is trained in a single source domain. Yet, the static DL model is likely to perform poorly in a continually evolving environment, requiring appropriate model updates. In an incremental learning setting, we would expect that well-trained static models are updated, following continually evolving target domain data -- e.g., additional lesions or structures of interest -- collected from different sites, without catastrophic forgetting. This, however, poses challenges, due to distribution shifts, additional structures not seen during the initial model training, and the absence of training data in a source domain. To address these challenges, in this work, we seek to progressively evolve an ``off-the-shelf" trained segmentation model to diverse datasets with additional anatomical categories in a unified manner. Specifically, we first propose a divergence-aware dual-flow module with balanced rigidity and plasticity branches to decouple old and new tasks, which is guided by continuous batch renormalization. Then, a complementary pseudo-label training scheme with self-entropy regularized momentum MixUp decay is developed for adaptive network optimization. We evaluated our framework on a brain tumor segmentation task with continually changing target domains -- i.e., new MRI scanners/modalities with incremental structures. Our framework was able to well retain the discriminability of previously learned structures, hence enabling the realistic life-long segmentation model extension along with the widespread accumulation of big medical data.
Gradient Descent (GD) has been proven effective in solving various matrix factorization problems. However, its optimization behavior with large initial values remains less understood. To address this gap, this paper presents a novel theoretical framework for examining the convergence trajectory of GD with a large initialization. The framework is grounded in signal-to-noise ratio concepts and inductive arguments. The results uncover an implicit incremental learning phenomenon in GD and offer a deeper understanding of its performance in large initialization scenarios.
In autonomous driving scenarios, current object detection models show strong performance when tested in clear weather. However, their performance deteriorates significantly when tested in degrading weather conditions. In addition, even when adapted to perform robustly in a sequence of different weather conditions, they are often unable to perform well in all of them and suffer from catastrophic forgetting. To efficiently mitigate forgetting, we propose Domain-Incremental Learning through Activation Matching (DILAM), which employs unsupervised feature alignment to adapt only the affine parameters of a clear weather pre-trained network to different weather conditions. We propose to store these affine parameters as a memory bank for each weather condition and plug-in their weather-specific parameters during driving (i.e. test time) when the respective weather conditions are encountered. Our memory bank is extremely lightweight, since affine parameters account for less than 2% of a typical object detector. Furthermore, contrary to previous domain-incremental learning approaches, we do not require the weather label when testing and propose to automatically infer the weather condition by a majority voting linear classifier.
Deep neural networks (DNNs) struggle to learn in dynamic environments since they rely on fixed datasets or stationary environments. Continual learning (CL) aims to address this limitation and enable DNNs to accumulate knowledge incrementally, similar to human learning. Inspired by how our brain consolidates memories, a powerful strategy in CL is replay, which involves training the DNN on a mixture of new and all seen classes. However, existing replay methods overlook two crucial aspects of biological replay: 1) the brain replays processed neural patterns instead of raw input, and 2) it prioritizes the replay of recently learned information rather than revisiting all past experiences. To address these differences, we propose SHARP, an efficient neuro-inspired CL method that leverages sparse dynamic connectivity and activation replay. Unlike other activation replay methods, which assume layers not subjected to replay have been pretrained and fixed, SHARP can continually update all layers. Also, SHARP is unique in that it only needs to replay few recently seen classes instead of all past classes. Our experiments on five datasets demonstrate that SHARP outperforms state-of-the-art replay methods in class incremental learning. Furthermore, we showcase SHARP's flexibility in a novel CL scenario where the boundaries between learning episodes are blurry. The SHARP code is available at \url{https://github.com/BurakGurbuz97/SHARP-Continual-Learning}.
Class-Incremental Learning (CIL) or continual learning is a desired capability in the real world, which requires a learning system to adapt to new tasks without forgetting former ones. While traditional CIL methods focus on visual information to grasp core features, recent advances in Vision-Language Models (VLM) have shown promising capabilities in learning generalizable representations with the aid of textual information. However, when continually trained with new classes, VLMs often suffer from catastrophic forgetting of former knowledge. Applying VLMs to CIL poses two major challenges: 1) how to adapt the model without forgetting; and 2) how to make full use of the multi-modal information. To this end, we propose PROjectiOn Fusion (PROOF) that enables VLMs to learn without forgetting. To handle the first challenge, we propose training task-specific projections based on the frozen image/text encoders. When facing new tasks, new projections are expanded and former projections are fixed, alleviating the forgetting of old concepts. For the second challenge, we propose the fusion module to better utilize the cross-modality information. By jointly adjusting visual and textual features, the model can capture semantic information with stronger representation ability. Extensive experiments on nine benchmark datasets validate PROOF achieves state-of-the-art performance.
Class-incremental learning (CIL) is a particularly challenging variant of continual learning, where the goal is to learn to discriminate between all classes presented in an incremental fashion. Existing approaches often suffer from excessive forgetting and imbalance of the scores assigned to classes that have not been seen together during training. In this study, we introduce a novel approach, Prediction Error-based Classification (PEC), which differs from traditional discriminative and generative classification paradigms. PEC computes a class score by measuring the prediction error of a model trained to replicate the outputs of a frozen random neural network on data from that class. The method can be interpreted as approximating a classification rule based on Gaussian Process posterior variance. PEC offers several practical advantages, including sample efficiency, ease of tuning, and effectiveness even when data are presented one class at a time. Our empirical results show that PEC performs strongly in single-pass-through-data CIL, outperforming other rehearsal-free baselines in all cases and rehearsal-based methods with moderate replay buffer size in most cases across multiple benchmarks.
Class-incremental learning is one of the most important settings for the study of Continual Learning, as it closely resembles real-world application scenarios. With constrained memory sizes, catastrophic forgetting arises as the number of classes/tasks increases. Studying continual learning in the video domain poses even more challenges, as video data contains a large number of frames, which places a higher burden on the replay memory. The current common practice is to sub-sample frames from the video stream and store them in the replay memory. In this paper, we propose SMILE a novel replay mechanism for effective video continual learning based on individual/single frames. Through extensive experimentation, we show that under extreme memory constraints, video diversity plays a more significant role than temporal information. Therefore, our method focuses on learning from a small number of frames that represent a large number of unique videos. On three representative video datasets, Kinetics, UCF101, and ActivityNet, the proposed method achieves state-of-the-art performance, outperforming the previous state-of-the-art by up to 21.49%.