Prompt tuning has emerged as an effective rehearsal-free technique for class-incremental learning (CIL) that learns a tiny set of task-specific parameters (or prompts) to instruct a pre-trained transformer to learn on a sequence of tasks. Albeit effective, prompt tuning methods do not lend well in the multi-label class incremental learning (MLCIL) scenario (where an image contains multiple foreground classes) due to the ambiguity in selecting the correct prompt(s) corresponding to different foreground objects belonging to multiple tasks. To circumvent this issue we propose to eliminate the prompt selection mechanism by maintaining task-specific pathways, which allow us to learn representations that do not interact with the ones from the other tasks. Since independent pathways in truly incremental scenarios will result in an explosion of computation due to the quadratically complex multi-head self-attention (MSA) operation in prompt tuning, we propose to reduce the original patch token embeddings into summarized tokens. Prompt tuning is then applied to these fewer summarized tokens to compute the final representation. Our proposed method Multi-Label class incremental learning via summarising pAtch tokeN Embeddings (MULTI-LANE) enables learning disentangled task-specific representations in MLCIL while ensuring fast inference. We conduct experiments in common benchmarks and demonstrate that our MULTI-LANE achieves a new state-of-the-art in MLCIL. Additionally, we show that MULTI-LANE is also competitive in the CIL setting. Source code available at https://github.com/tdemin16/multi-lane
Continual learning (CL) remains a significant challenge for deep neural networks, as it is prone to forgetting previously acquired knowledge. Several approaches have been proposed in the literature, such as experience rehearsal, regularization, and parameter isolation, to address this problem. Although almost zero forgetting can be achieved in task-incremental learning, class-incremental learning remains highly challenging due to the problem of inter-task class separation. Limited access to previous task data makes it difficult to discriminate between classes of current and previous tasks. To address this issue, we propose `Attention-Guided Incremental Learning' (AGILE), a novel rehearsal-based CL approach that incorporates compact task attention to effectively reduce interference between tasks. AGILE utilizes lightweight, learnable task projection vectors to transform the latent representations of a shared task attention module toward task distribution. Through extensive empirical evaluation, we show that AGILE significantly improves generalization performance by mitigating task interference and outperforming rehearsal-based approaches in several CL scenarios. Furthermore, AGILE can scale well to a large number of tasks with minimal overhead while remaining well-calibrated with reduced task-recency bias.
Class incremental learning consists in training discriminative models to classify an increasing number of classes over time. However, doing so using only the newly added class data leads to the known problem of catastrophic forgetting of the previous classes. Recently, dynamic deep learning architectures have been shown to exhibit a better stability-plasticity trade-off by dynamically adding new feature extractors to the model in order to learn new classes followed by a compression step to scale the model back to its original size, thus avoiding a growing number of parameters. In this context, we propose a new algorithm that enhances the compression of previous class knowledge by cutting and mixing patches of previous class samples with the new images during compression using our Rehearsal-CutMix method. We show that this new data augmentation reduces catastrophic forgetting by specifically targeting past class information and improving its compression. Extensive experiments performed on the CIFAR and ImageNet datasets under diverse incremental learning evaluation protocols demonstrate that our approach consistently outperforms the state-of-the-art . The code will be made available upon publication of our work.
Few-Shot Class-Incremental Learning presents an extension of the Class Incremental Learning problem where a model is faced with the problem of data scarcity while addressing the catastrophic forgetting problem. This problem remains an open problem because all recent works are built upon the convolutional neural networks performing sub-optimally compared to the transformer approaches. Our paper presents Robust Transformer Approach built upon the Compact Convolution Transformer. The issue of overfitting due to few samples is overcome with the notion of the stochastic classifier, where the classifier's weights are sampled from a distribution with mean and variance vectors, thus increasing the likelihood of correct classifications, and the batch-norm layer to stabilize the training process. The issue of CF is dealt with the idea of delta parameters, small task-specific trainable parameters while keeping the backbone networks frozen. A non-parametric approach is developed to infer the delta parameters for the model's predictions. The prototype rectification approach is applied to avoid biased prototype calculations due to the issue of data scarcity. The advantage of ROBUSTA is demonstrated through a series of experiments in the benchmark problems where it is capable of outperforming prior arts with big margins without any data augmentation protocols.
Efforts to overcome catastrophic forgetting have primarily centered around developing more effective Continual Learning (CL) methods. In contrast, less attention was devoted to analyzing the role of network architecture design (e.g., network depth, width, and components) in contributing to CL. This paper seeks to bridge this gap between network architecture design and CL, and to present a holistic study on the impact of network architectures on CL. This work considers architecture design at the network scaling level, i.e., width and depth, and also at the network components, i.e., skip connections, global pooling layers, and down-sampling. In both cases, we first derive insights through systematically exploring how architectural designs affect CL. Then, grounded in these insights, we craft a specialized search space for CL and further propose a simple yet effective ArchCraft method to steer a CL-friendly architecture, namely, this method recrafts AlexNet/ResNet into AlexAC/ResAC. Experimental validation across various CL settings and scenarios demonstrates that improved architectures are parameter-efficient, achieving state-of-the-art performance of CL while being 86%, 61%, and 97% more compact in terms of parameters than the naive CL architecture in Task IL and Class IL. Code is available at https://github.com/byyx666/ArchCraft.
Continual learning, the ability to acquire knowledge from new data while retaining previously learned information, is a fundamental challenge in machine learning. Various approaches, including memory replay, knowledge distillation, model regularization, and dynamic network expansion, have been proposed to address this issue. Thus far, dynamic network expansion methods have achieved state-of-the-art performance at the cost of incurring significant computational overhead. This is due to the need for additional model buffers, which makes it less feasible in resource-constrained settings, particularly in the medical domain. To overcome this challenge, we propose Dynamic Model Merging, DynaMMo, a method that merges multiple networks at different stages of model training to achieve better computational efficiency. Specifically, we employ lightweight learnable modules for each task and combine them into a unified model to minimize computational overhead. DynaMMo achieves this without compromising performance, offering a cost-effective solution for continual learning in medical applications. We evaluate DynaMMo on three publicly available datasets, demonstrating its effectiveness compared to existing approaches. DynaMMo offers around 10-fold reduction in GFLOPS with a small drop of 2.76 in average accuracy when compared to state-of-the-art dynamic-based approaches. The code implementation of this work will be available upon the acceptance of this work at https://github.com/BioMedIA-MBZUAI/DynaMMo.
Few-shot class-incremental learning (FSCIL) aims to adapt the model to new classes from very few data (5 samples) without forgetting the previously learned classes. Recent works in many-shot CIL (MSCIL) (using all available training data) exploited pre-trained models to reduce forgetting and achieve better plasticity. In a similar fashion, we use ViT models pre-trained on large-scale datasets for few-shot settings, which face the critical issue of low plasticity. FSCIL methods start with a many-shot first task to learn a very good feature extractor and then move to the few-shot setting from the second task onwards. While the focus of most recent studies is on how to learn the many-shot first task so that the model generalizes to all future few-shot tasks, we explore in this work how to better model the few-shot data using pre-trained models, irrespective of how the first task is trained. Inspired by recent works in MSCIL, we explore how using higher-order feature statistics can influence the classification of few-shot classes. We identify the main challenge of obtaining a good covariance matrix from few-shot data and propose to calibrate the covariance matrix for new classes based on semantic similarity to the many-shot base classes. Using the calibrated feature statistics in combination with existing methods significantly improves few-shot continual classification on several FSCIL benchmarks. Code is available at https://github.com/dipamgoswami/FSCIL-Calibration.
Few-Shot Class Incremental Learning (FSCIL) is a task that requires a model to learn new classes incrementally without forgetting when only a few samples for each class are given. FSCIL encounters two significant challenges: catastrophic forgetting and overfitting, and these challenges have driven prior studies to primarily rely on shallow models, such as ResNet-18. Even though their limited capacity can mitigate both forgetting and overfitting issues, it leads to inadequate knowledge transfer during few-shot incremental sessions. In this paper, we argue that large models such as vision and language transformers pre-trained on large datasets can be excellent few-shot incremental learners. To this end, we propose a novel FSCIL framework called PriViLege, Pre-trained Vision and Language transformers with prompting functions and knowledge distillation. Our framework effectively addresses the challenges of catastrophic forgetting and overfitting in large models through new pre-trained knowledge tuning (PKT) and two losses: entropy-based divergence loss and semantic knowledge distillation loss. Experimental results show that the proposed PriViLege significantly outperforms the existing state-of-the-art methods with a large margin, e.g., +9.38% in CUB200, +20.58% in CIFAR-100, and +13.36% in miniImageNet. Our implementation code is available at https://github.com/KHU-AGI/PriViLege.
Class-incremental learning (CIL) aims to enable models to continuously learn new classes while overcoming catastrophic forgetting. The introduction of pre-trained models has brought new tuning paradigms to CIL. In this paper, we revisit different parameter-efficient tuning (PET) methods within the context of continual learning. We observe that adapter tuning demonstrates superiority over prompt-based methods, even without parameter expansion in each learning session. Motivated by this, we propose incrementally tuning the shared adapter without imposing parameter update constraints, enhancing the learning capacity of the backbone. Additionally, we employ feature sampling from stored prototypes to retrain a unified classifier, further improving its performance. We estimate the semantic shift of old prototypes without access to past samples and update stored prototypes session by session. Our proposed method eliminates model expansion and avoids retaining any image samples. It surpasses previous pre-trained model-based CIL methods and demonstrates remarkable continual learning capabilities. Experimental results on five CIL benchmarks validate the effectiveness of our approach, achieving state-of-the-art (SOTA) performance.
Few-Shot Class-Incremental Learning (FSCIL) introduces a paradigm in which the problem space expands with limited data. FSCIL methods inherently face the challenge of catastrophic forgetting as data arrives incrementally, making models susceptible to overwriting previously acquired knowledge. Moreover, given the scarcity of labeled samples available at any given time, models may be prone to overfitting and find it challenging to strike a balance between extensive pretraining and the limited incremental data. To address these challenges, we propose the OrCo framework built on two core principles: features' orthogonality in the representation space, and contrastive learning. In particular, we improve the generalization of the embedding space by employing a combination of supervised and self-supervised contrastive losses during the pretraining phase. Additionally, we introduce OrCo loss to address challenges arising from data limitations during incremental sessions. Through feature space perturbations and orthogonality between classes, the OrCo loss maximizes margins and reserves space for the following incremental data. This, in turn, ensures the accommodation of incoming classes in the feature space without compromising previously acquired knowledge. Our experimental results showcase state-of-the-art performance across three benchmark datasets, including mini-ImageNet, CIFAR100, and CUB datasets. Code is available at https://github.com/noorahmedds/OrCo
In class-incremental learning (CIL) scenarios, the phenomenon of catastrophic forgetting caused by the classifier's bias towards the current task has long posed a significant challenge. It is mainly caused by the characteristic of discriminative models. With the growing popularity of the generative multi-modal models, we would explore replacing discriminative models with generative ones for CIL. However, transitioning from discriminative to generative models requires addressing two key challenges. The primary challenge lies in transferring the generated textual information into the classification of distinct categories. Additionally, it requires formulating the task of CIL within a generative framework. To this end, we propose a novel generative multi-modal model (GMM) framework for class-incremental learning. Our approach directly generates labels for images using an adapted generative model. After obtaining the detailed text, we use a text encoder to extract text features and employ feature matching to determine the most similar label as the classification prediction. In the conventional CIL settings, we achieve significantly better results in long-sequence task scenarios. Under the Few-shot CIL setting, we have improved by at least 14\% accuracy over all the current state-of-the-art methods with significantly less forgetting. Our code is available at \url{https://github.com/DoubleClass/GMM}.
Class-incremental learning (CIL) under an exemplar-free constraint has presented a significant challenge. Existing methods adhering to this constraint are prone to catastrophic forgetting, far more so than replay-based techniques that retain access to past samples. In this paper, to solve the exemplar-free CIL problem, we propose a Dual-Stream Analytic Learning (DS-AL) approach. The DS-AL contains a main stream offering an analytical (i.e., closed-form) linear solution, and a compensation stream improving the inherent under-fitting limitation due to adopting linear mapping. The main stream redefines the CIL problem into a Concatenated Recursive Least Squares (C-RLS) task, allowing an equivalence between the CIL and its joint-learning counterpart. The compensation stream is governed by a Dual-Activation Compensation (DAC) module. This module re-activates the embedding with a different activation function from the main stream one, and seeks fitting compensation by projecting the embedding to the null space of the main stream's linear mapping. Empirical results demonstrate that the DS-AL, despite being an exemplar-free technique, delivers performance comparable with or better than that of replay-based methods across various datasets, including CIFAR-100, ImageNet-100 and ImageNet-Full. Additionally, the C-RLS' equivalent property allows the DS-AL to execute CIL in a phase-invariant manner. This is evidenced by a never-before-seen 500-phase CIL ImageNet task, which performs on a level identical to a 5-phase one. Our codes are available at https://github.com/ZHUANGHP/Analytic-continual-learning.
Online Class Incremental Learning (OCIL) aims to train models incrementally, where data arrive in mini-batches, and previous data are not accessible. A major challenge in OCIL is Catastrophic Forgetting, i.e., the loss of previously learned knowledge. Among existing baselines, replay-based methods show competitive results but requires extra memory for storing exemplars, while exemplar-free (i.e., data need not be stored for replay in production) methods are resource-friendly but often lack accuracy. In this paper, we propose an exemplar-free approach--Forward-only Online Analytic Learning (F-OAL). Unlike traditional methods, F-OAL does not rely on back-propagation and is forward-only, significantly reducing memory usage and computational time. Cooperating with a pre-trained frozen encoder with Feature Fusion, F-OAL only needs to update a linear classifier by recursive least square. This approach simultaneously achieves high accuracy and low resource consumption. Extensive experiments on benchmark datasets demonstrate F-OAL's robust performance in OCIL scenarios. Code is available at https://github.com/liuyuchen-cz/F-OAL.
Class incremental learning (CIL) trains a network on sequential tasks with separated categories in each task but suffers from catastrophic forgetting, where models quickly lose previously learned knowledge when acquiring new tasks. The generalized CIL (GCIL) aims to address the CIL problem in a more real-world scenario, where incoming data have mixed data categories and unknown sample size distribution. Existing attempts for the GCIL either have poor performance or invade data privacy by saving exemplars. In this paper, we propose a new exemplar-free GCIL technique named generalized analytic continual learning (GACL). The GACL adopts analytic learning (a gradient-free training technique) and delivers an analytical (i.e., closed-form) solution to the GCIL scenario. This solution is derived via decomposing the incoming data into exposed and unexposed classes, thereby attaining a weight-invariant property, a rare yet valuable property supporting an equivalence between incremental learning and its joint training. Such an equivalence is crucial in GCIL settings as data distributions among different tasks no longer pose challenges to adopting our GACL. Theoretically, this equivalence property is validated through matrix analysis tools. Empirically, we conduct extensive experiments where, compared with existing GCIL methods, our GACL exhibits a consistently leading performance across various datasets and GCIL settings. Source code is available at https://github.com/CHEN-YIZHU/GACL.
Class Incremental Learning (CIL) constitutes a pivotal subfield within continual learning, aimed at enabling models to progressively learn new classification tasks while retaining knowledge obtained from prior tasks. Although previous studies have predominantly focused on backward compatible approaches to mitigate catastrophic forgetting, recent investigations have introduced forward compatible methods to enhance performance on novel tasks and complement existing backward compatible methods. In this study, we introduce an effective-Rank based Feature Richness enhancement (RFR) method, designed for improving forward compatibility. Specifically, this method increases the effective rank of representations during the base session, thereby facilitating the incorporation of more informative features pertinent to unseen novel tasks. Consequently, RFR achieves dual objectives in backward and forward compatibility: minimizing feature extractor modifications and enhancing novel task performance, respectively. To validate the efficacy of our approach, we establish a theoretical connection between effective rank and the Shannon entropy of representations. Subsequently, we conduct comprehensive experiments by integrating RFR into eleven well-known CIL methods. Our results demonstrate the effectiveness of our approach in enhancing novel-task performance while mitigating catastrophic forgetting. Furthermore, our method notably improves the average incremental accuracy across all eleven cases examined.
Class incremental learning (CIL) in an online continual learning setting strives to acquire knowledge on a series of novel classes from a data stream, using each data point only once for training. This is more realistic compared to offline modes, where it is assumed that all data from novel class(es) is readily available. Current online CIL approaches store a subset of the previous data which creates heavy overhead costs in terms of both memory and computation, as well as privacy issues. In this paper, we propose a new online CIL approach called FOCIL. It fine-tunes the main architecture continually by training a randomly pruned sparse subnetwork for each task. Then, it freezes the trained connections to prevent forgetting. FOCIL also determines the sparsity level and learning rate per task adaptively and ensures (almost) zero forgetting across all tasks without storing any replay data. Experimental results on 10-Task CIFAR100, 20-Task CIFAR100, and 100-Task TinyImagenet, demonstrate that our method outperforms the SOTA by a large margin. The code is publicly available at https://github.com/muratonuryildirim/FOCIL.
Federated Class-Incremental Learning (FCIL) is an underexplored yet pivotal issue, involving the dynamic addition of new classes in the context of federated learning. In this field, Data-Free Knowledge Transfer (DFKT) plays a crucial role in addressing catastrophic forgetting and data privacy problems. However, prior approaches lack the crucial synergy between DFKT and the model training phases, causing DFKT to encounter difficulties in generating high-quality data from a non-anchored latent space of the old task model. In this paper, we introduce LANDER (Label Text Centered Data-Free Knowledge Transfer) to address this issue by utilizing label text embeddings (LTE) produced by pretrained language models. Specifically, during the model training phase, our approach treats LTE as anchor points and constrains the feature embeddings of corresponding training samples around them, enriching the surrounding area with more meaningful information. In the DFKT phase, by using these LTE anchors, LANDER can synthesize more meaningful samples, thereby effectively addressing the forgetting problem. Additionally, instead of tightly constraining embeddings toward the anchor, the Bounding Loss is introduced to encourage sample embeddings to remain flexible within a defined radius. This approach preserves the natural differences in sample embeddings and mitigates the embedding overlap caused by heterogeneous federated settings. Extensive experiments conducted on CIFAR100, Tiny-ImageNet, and ImageNet demonstrate that LANDER significantly outperforms previous methods and achieves state-of-the-art performance in FCIL. The code is available at https://github.com/tmtuan1307/lander.
The proliferation of Few-Shot Class Incremental Learning (FSCIL) methodologies has highlighted the critical challenge of maintaining robust anti-amnesia capabilities in FSCIL learners. In this paper, we present a novel conceptualization of anti-amnesia in terms of mathematical generalization, leveraging the Neural Tangent Kernel (NTK) perspective. Our method focuses on two key aspects: ensuring optimal NTK convergence and minimizing NTK-related generalization loss, which serve as the theoretical foundation for cross-task generalization. To achieve global NTK convergence, we introduce a principled meta-learning mechanism that guides optimization within an expanded network architecture. Concurrently, to reduce the NTK-related generalization loss, we systematically optimize its constituent factors. Specifically, we initiate self-supervised pre-training on the base session to enhance NTK-related generalization potential. These self-supervised weights are then carefully refined through curricular alignment, followed by the application of dual NTK regularization tailored specifically for both convolutional and linear layers. Through the combined effects of these measures, our network acquires robust NTK properties, ensuring optimal convergence and stability of the NTK matrix and minimizing the NTK-related generalization loss, significantly enhancing its theoretical generalization. On popular FSCIL benchmark datasets, our NTK-FSCIL surpasses contemporary state-of-the-art approaches, elevating end-session accuracy by 2.9\% to 9.3\%.
Class-Incremental Learning (CIL) requires a learning system to continually learn new classes without forgetting. Despite the strong performance of Pre-Trained Models (PTMs) in CIL, a critical issue persists: learning new classes often results in the overwriting of old ones. Excessive modification of the network causes forgetting, while minimal adjustments lead to an inadequate fit for new classes. As a result, it is desired to figure out a way of efficient model updating without harming former knowledge. In this paper, we propose ExpAndable Subspace Ensemble (EASE) for PTM-based CIL. To enable model updating without conflict, we train a distinct lightweight adapter module for each new task, aiming to create task-specific subspaces. These adapters span a high-dimensional feature space, enabling joint decision-making across multiple subspaces. As data evolves, the expanding subspaces render the old class classifiers incompatible with new-stage spaces. Correspondingly, we design a semantic-guided prototype complement strategy that synthesizes old classes' new features without using any old class instance. Extensive experiments on seven benchmark datasets verify EASE's state-of-the-art performance. Code is available at: https://github.com/sun-hailong/CVPR24-Ease
This paper explores the problem of continual learning (CL) of vision-language models (VLMs) in open domains, where the models need to perform continual updating and inference on a streaming of datasets from diverse seen and unseen domains with novel classes. Such a capability is crucial for various applications in open environments, e.g., AI assistants, autonomous driving systems, and robotics. Current CL studies mostly focus on closed-set scenarios in a single domain with known classes. Large pre-trained VLMs like CLIP have demonstrated superior zero-shot recognition ability, and a number of recent studies leverage this ability to mitigate catastrophic forgetting in CL, but they focus on closed-set CL in a single domain dataset. Open-domain CL of large VLMs is significantly more challenging due to 1) large class correlations and domain gaps across the datasets and 2) the forgetting of zero-shot knowledge in the pre-trained VLMs in addition to the knowledge learned from the newly adapted datasets. In this work we introduce a novel approach, termed CoLeCLIP, that learns an open-domain CL model based on CLIP. It addresses these challenges by a joint learning of a set of task prompts and a cross-domain class vocabulary. Extensive experiments on 11 domain datasets show that CoLeCLIP outperforms state-of-the-art methods for open-domain CL under both task- and class-incremental learning settings.