This study focuses on incremental learning for image classification, exploring how to reduce catastrophic forgetting of all learned knowledge when access to old data is restricted due to memory or privacy constraints. The challenge of incremental learning lies in achieving an optimal balance between plasticity, the ability to learn new knowledge, and stability, the ability to retain old knowledge. Based on whether the task identifier (task-ID) of an image can be obtained during the test stage, incremental learning for image classifcation is divided into two main paradigms, which are task incremental learning (TIL) and class incremental learning (CIL). The TIL paradigm has access to the task-ID, allowing it to use multiple task-specific classification heads selected based on the task-ID. Consequently, in CIL, where the task-ID is unavailable, TIL methods must predict the task-ID to extend their application to the CIL paradigm. Our previous method for TIL adds task-specific batch normalization and classification heads incrementally. This work extends the method by predicting task-ID through an "unknown" class added to each classification head. The head with the lowest "unknown" probability is selected, enabling task-ID prediction and making the method applicable to CIL. The task-specific batch normalization (BN) modules effectively adjust the distribution of output feature maps across different tasks, enhancing the model's plasticity.Moreover, since BN has much fewer parameters compared to convolutional kernels, by only modifying the BN layers as new tasks arrive, the model can effectively manage parameter growth while ensuring stability across tasks. The innovation of this study lies in the first-time introduction of task-specific BN into CIL and verifying the feasibility of extending TIL methods to CIL through task-ID prediction with state-of-the-art performance on multiple datasets.
Despite the outstanding performance in many individual tasks, deep neural networks suffer from catastrophic forgetting when learning from continuous data streams in real-world scenarios. Current Non-Exemplar Class-Incremental Learning (NECIL) methods mitigate forgetting by storing a single prototype per class, which serves to inject previous information when sequentially learning new classes. However, these stored prototypes or their augmented variants often fail to simultaneously capture spatial distribution diversity and precision needed for representing old classes. Moreover, as the model acquires new knowledge, these prototypes gradually become outdated, making them less effective. To overcome these limitations, we propose a more efficient NECIL method that replaces prototypes with synthesized retrospective features for old classes. Specifically, we model each old class's feature space using a multivariate Gaussian distribution and generate deep representations by sampling from high-likelihood regions. Additionally, we introduce a similarity-based feature compensation mechanism that integrates generated old class features with similar new class features to synthesize robust retrospective representations. These retrospective features are then incorporated into our incremental learning framework to preserve the decision boundaries of previous classes while learning new ones. Extensive experiments on CIFAR-100, TinyImageNet, and ImageNet-Subset demonstrate that our method significantly improves the efficiency of non-exemplar class-incremental learning and achieves state-of-the-art performance.
Few-shot Class Incremental Learning (FSCIL) presents a challenging yet realistic scenario, which requires the model to continually learn new classes with limited labeled data (i.e., incremental sessions) while retaining knowledge of previously learned base classes (i.e., base sessions). Due to the limited data in incremental sessions, models are prone to overfitting new classes and suffering catastrophic forgetting of base classes. To tackle these issues, recent advancements resort to prototype-based approaches to constrain the base class distribution and learn discriminative representations of new classes. Despite the progress, the limited data issue still induces ill-divided feature space, leading the model to confuse the new class with old classes or fail to facilitate good separation among new classes. In this paper, we aim to mitigate these issues by directly constraining the span of each class distribution from a covariance perspective. In detail, we propose a simple yet effective covariance constraint loss to force the model to learn each class distribution with the same covariance matrix. In addition, we propose a perturbation approach to perturb the few-shot training samples in the feature space, which encourages the samples to be away from the weighted distribution of other classes. Regarding perturbed samples as new class data, the classifier is forced to establish explicit boundaries between each new class and the existing ones. Our approach is easy to integrate into existing FSCIL approaches to boost performance. Experiments on three benchmarks validate the effectiveness of our approach, achieving a new state-of-the-art performance of FSCIL.
Deep neural networks (DNNS) excel at learning from static datasets but struggle with continual learning, where data arrives sequentially. Catastrophic forgetting, the phenomenon of forgetting previously learned knowledge, is a primary challenge. This paper introduces EXponentially Averaged Class-wise Feature Significance (EXACFS) to mitigate this issue in the class incremental learning (CIL) setting. By estimating the significance of model features for each learned class using loss gradients, gradually aging the significance through the incremental tasks and preserving the significant features through a distillation loss, EXACFS effectively balances remembering old knowledge (stability) and learning new knowledge (plasticity). Extensive experiments on CIFAR-100 and ImageNet-100 demonstrate EXACFS's superior performance in preserving stability while acquiring plasticity.
Deep neural networks can struggle to learn continually in the face of non-stationarity. This phenomenon is known as loss of plasticity. In this paper, we identify underlying principles that lead to plastic algorithms. In particular, we provide theoretical results showing that linear function approximation, as well as a special case of deep linear networks, do not suffer from loss of plasticity. We then propose deep Fourier features, which are the concatenation of a sine and cosine in every layer, and we show that this combination provides a dynamic balance between the trainability obtained through linearity and the effectiveness obtained through the nonlinearity of neural networks. Deep networks composed entirely of deep Fourier features are highly trainable and sustain their trainability over the course of learning. Our empirical results show that continual learning performance can be drastically improved by replacing ReLU activations with deep Fourier features. These results hold for different continual learning scenarios (e.g., label noise, class incremental learning, pixel permutations) on all major supervised learning datasets used for continual learning research, such as CIFAR10, CIFAR100, and tiny-ImageNet.
In class-incremental learning (class-IL), models must classify all previously seen classes at test time without task-IDs, leading to task confusion. Despite being a key challenge, task confusion lacks a theoretical understanding. We present a novel mathematical framework for class-IL and prove the Infeasibility Theorem, showing optimal class-IL is impossible with discriminative modeling due to task confusion. However, we establish the Feasibility Theorem, demonstrating that generative modeling can achieve optimal class-IL by overcoming task confusion. We then assess popular class-IL strategies, including regularization, bias-correction, replay, and generative classifier, using our framework. Our analysis suggests that adopting generative modeling, either for generative replay or direct classification (generative classifier), is essential for optimal class-IL.
Model merging has emerged as a crucial technique in Deep Learning, enabling the integration of multiple models into a unified system while preserving performance and scalability. In this respect, the compositional properties of low-rank adaptation techniques (e.g., LoRA) have proven beneficial, as simple averaging LoRA modules yields a single model that mostly integrates the capabilities of all individual modules. Building on LoRA, we take a step further by imposing that the merged model matches the responses of all learned modules. Solving this objective in closed form yields an indeterminate system with A and B as unknown variables, indicating the existence of infinitely many closed-form solutions. To address this challenge, we introduce LoRM, an alternating optimization strategy that trains one LoRA matrix at a time. This allows solving for each unknown variable individually, thus finding a unique solution. We apply our proposed methodology to Federated Class-Incremental Learning (FCIL), ensuring alignment of model responses both between clients and across tasks. Our method demonstrates state-of-the-art performance across a range of FCIL scenarios.
Incremental learning (IL) aims to acquire new knowledge from current tasks while retaining knowledge learned from previous tasks. Replay-based IL methods store a set of exemplars from previous tasks in a buffer and replay them when learning new tasks. However, there is usually a size-limited buffer that cannot store adequate real exemplars to retain the knowledge of previous tasks. In contrast, data distillation (DD) can reduce the exemplar buffer's size, by condensing a large real dataset into a much smaller set of more information-compact synthetic exemplars. Nevertheless, DD's performance gain on IL quickly vanishes as the number of synthetic exemplars grows. To overcome the weaknesses of real-data and synthetic-data buffers, we instead optimize a hybrid memory including both types of data. Specifically, we propose an innovative modification to DD that distills synthetic data from a sliding window of checkpoints in history (rather than checkpoints on multiple training trajectories). Conditioned on the synthetic data, we then optimize the selection of real exemplars to provide complementary improvement to the DD objective. The optimized hybrid memory combines the strengths of synthetic and real exemplars, effectively mitigating catastrophic forgetting in Class IL (CIL) when the buffer size for exemplars is limited. Notably, our method can be seamlessly integrated into most existing replay-based CIL models. Extensive experiments across multiple benchmarks demonstrate that our method significantly outperforms existing replay-based baselines.
Time series classification underpins critical applications such as healthcare diagnostics and gesture-driven interactive systems in multimedia scenarios. However, time series class-incremental learning (TSCIL) faces two major challenges: catastrophic forgetting and intra-class variations. Catastrophic forgetting occurs because gradient-based parameter update strategies inevitably erase past knowledge. And unlike images, time series data exhibits subject-specific patterns, also known as intra-class variations, which refer to differences in patterns observed within the same class. While exemplar-based methods fail to cover diverse variation with limited samples, existing exemplar-free methods lack explicit mechanisms to handle intra-class variations. To address these two challenges, we propose TS-ACL, which leverages a gradient-free closed-form solution to avoid the catastrophic forgetting problem inherent in gradient-based optimization methods while simultaneously learning global distributions to resolve intra-class variations. Additionally, it provides privacy protection and efficiency. Extensive experiments on five benchmark datasets covering various sensor modalities and tasks demonstrate that TS-ACL achieves performance close to joint training on four datasets, outperforming existing methods and establishing a new state-of-the-art (SOTA) for TSCIL.
Class-incremental learning (CIL) aims to continually learn a sequence of tasks, with each task consisting of a set of unique classes. Graph CIL (GCIL) follows the same setting but needs to deal with graph tasks (e.g., node classification in a graph). The key characteristic of CIL lies in the absence of task identifiers (IDs) during inference, which causes a significant challenge in separating classes from different tasks (i.e., inter-task class separation). Being able to accurately predict the task IDs can help address this issue, but it is a challenging problem. In this paper, we show theoretically that accurate task ID prediction on graph data can be achieved by a Laplacian smoothing-based graph task profiling approach, in which each graph task is modeled by a task prototype based on Laplacian smoothing over the graph. It guarantees that the task prototypes of the same graph task are nearly the same with a large smoothing step, while those of different tasks are distinct due to differences in graph structure and node attributes. Further, to avoid the catastrophic forgetting of the knowledge learned in previous graph tasks, we propose a novel graph prompting approach for GCIL which learns a small discriminative graph prompt for each task, essentially resulting in a separate classification model for each task. The prompt learning requires the training of a single graph neural network (GNN) only once on the first task, and no data replay is required thereafter, thereby obtaining a GCIL model being both replay-free and forget-free. Extensive experiments on four GCIL benchmarks show that i) our task prototype-based method can achieve 100% task ID prediction accuracy on all four datasets, ii) our GCIL model significantly outperforms state-of-the-art competing methods by at least 18% in average CIL accuracy, and iii) our model is fully free of forgetting on the four datasets.
Recent advances in deep learning for processing point clouds hold increased interest in Few-Shot Class Incremental Learning (FSCIL) for 3D computer vision. This paper introduces a new method to tackle the Few-Shot Continual Incremental Learning (FSCIL) problem in 3D point cloud environments. We leverage a foundational 3D model trained extensively on point cloud data. Drawing from recent improvements in foundation models, known for their ability to work well across different tasks, we propose a novel strategy that does not require additional training to adapt to new tasks. Our approach uses a dual cache system: first, it uses previous test samples based on how confident the model was in its predictions to prevent forgetting, and second, it includes a small number of new task samples to prevent overfitting. This dynamic adaptation ensures strong performance across different learning tasks without needing lots of fine-tuning. We tested our approach on datasets like ModelNet, ShapeNet, ScanObjectNN, and CO3D, showing that it outperforms other FSCIL methods and demonstrating its effectiveness and versatility. The code is available at \url{https://github.com/ahmadisahar/ACCV_FCIL3D}.
Pre-trained models have demonstrated impressive generalization capabilities, yet they remain vulnerable to catastrophic forgetting when incrementally trained on new tasks. Existing architecture-based strategies encounter two primary challenges: 1) Integrating a pre-trained network with a trainable sub-network complicates the delicate balance between learning plasticity and memory stability across evolving tasks during learning. 2) The absence of robust interconnections between pre-trained networks and various sub-networks limits the effective retrieval of pertinent information during inference. In this study, we introduce the Artsy, inspired by the activation mechanisms of silent synapses via spike-timing-dependent plasticity observed in mature brains, to enhance the continual learning capabilities of pre-trained models. The Artsy integrates two key components: During training, the Artsy mimics mature brain dynamics by maintaining memory stability for previously learned knowledge within the pre-trained network while simultaneously promoting learning plasticity in task-specific sub-networks. During inference, artificial silent and functional synapses are utilized to establish precise connections between the pre-synaptic neurons in the pre-trained network and the post-synaptic neurons in the sub-networks, facilitated through synaptic consolidation, thereby enabling effective extraction of relevant information from test samples. Comprehensive experimental evaluations reveal that our model significantly outperforms conventional methods on class-incremental learning tasks, while also providing enhanced biological interpretability for architecture-based approaches. Moreover, we propose that the Artsy offers a promising avenue for simulating biological synaptic mechanisms, potentially advancing our understanding of neural plasticity in both artificial and biological systems.
Aiming to incrementally learn new classes with only few samples while preserving the knowledge of base (old) classes, few-shot class-incremental learning (FSCIL) faces several challenges, such as overfitting and catastrophic forgetting. Such a challenging problem is often tackled by fixing a feature extractor trained on base classes to reduce the adverse effects of overfitting and forgetting. Under such formulation, our primary focus is representation learning on base classes to tackle the unique challenge of FSCIL: simultaneously achieving the transferability and the discriminability of the learned representation. Building upon the recent efforts for enhancing transferability, such as promoting the spread of features, we find that trying to secure the spread of features within a more confined feature space enables the learned representation to strike a better balance between transferability and discriminability. Thus, in stark contrast to prior beliefs that the inter-class distance should be maximized, we claim that the closer different classes are, the better for FSCIL. The empirical results and analysis from the perspective of information bottleneck theory justify our simple yet seemingly counter-intuitive representation learning method, raising research questions and suggesting alternative research directions. The code is available at https://github.com/JungHunOh/CLOSER_ECCV2024.
Model fairness is becoming important in class-incremental learning for Trustworthy AI. While accuracy has been a central focus in class-incremental learning, fairness has been relatively understudied. However, naively using all the samples of the current task for training results in unfair catastrophic forgetting for certain sensitive groups including classes. We theoretically analyze that forgetting occurs if the average gradient vector of the current task data is in an "opposite direction" compared to the average gradient vector of a sensitive group, which means their inner products are negative. We then propose a fair class-incremental learning framework that adjusts the training weights of current task samples to change the direction of the average gradient vector and thus reduce the forgetting of underperforming groups and achieve fairness. For various group fairness measures, we formulate optimization problems to minimize the overall losses of sensitive groups while minimizing the disparities among them. We also show the problems can be solved with linear programming and propose an efficient Fairness-aware Sample Weighting (FSW) algorithm. Experiments show that FSW achieves better accuracy-fairness tradeoff results than state-of-the-art approaches on real datasets.
Exemplar-Free Class Incremental Learning (EFCIL) tackles the problem of training a model on a sequence of tasks without access to past data. Existing state-of-the-art methods represent classes as Gaussian distributions in the feature extractor's latent space, enabling Bayes classification or training the classifier by replaying pseudo features. However, we identify two critical issues that compromise their efficacy when the feature extractor is updated on incremental tasks. First, they do not consider that classes' covariance matrices change and must be adapted after each task. Second, they are susceptible to a task-recency bias caused by dimensionality collapse occurring during training. In this work, we propose AdaGauss -- a novel method that adapts covariance matrices from task to task and mitigates the task-recency bias owing to the additional anti-collapse loss function. AdaGauss yields state-of-the-art results on popular EFCIL benchmarks and datasets when training from scratch or starting from a pre-trained backbone. The code is available at: https://github.com/grypesc/AdaGauss.
Continual learning (CL) adapt the deep learning scenarios with timely updated datasets. However, existing CL models suffer from the catastrophic forgetting issue, where new knowledge replaces past learning. In this paper, we propose Continual Learning with Task Specialists (CLTS) to address the issues of catastrophic forgetting and limited labelled data in real-world datasets by performing class incremental learning of the incoming stream of data. The model consists of Task Specialists (T S) and Task Predictor (T P ) with pre-trained Stable Diffusion (SD) module. Here, we introduce a new specialist to handle a new task sequence and each T S has three blocks; i) a variational autoencoder (V AE) to learn the task distribution in a low dimensional latent space, ii) a K-Means block to perform data clustering and iii) Bootstrapping Language-Image Pre-training (BLIP ) model to generate a small batch of captions from the input data. These captions are fed as input to the pre-trained stable diffusion model (SD) for the generation of task samples. The proposed model does not store any task samples for replay, instead uses generated samples from SD to train the T P module. A comparison study with four SOTA models conducted on three real-world datasets shows that the proposed model outperforms all the selected baselines
Non-exemplar class Incremental Learning (NECIL) enables models to continuously acquire new classes without retraining from scratch and storing old task exemplars, addressing privacy and storage issues. However, the absence of data from earlier tasks exacerbates the challenge of catastrophic forgetting in NECIL. In this paper, we propose a novel framework called Dynamic Integration of task-specific Adapters (DIA), which comprises two key components: Task-Specific Adapter Integration (TSAI) and Patch-Level Model Alignment. TSAI boosts compositionality through a patch-level adapter integration strategy, which provides a more flexible compositional solution while maintaining low computation costs. Patch-Level Model Alignment maintains feature consistency and accurate decision boundaries via two specialized mechanisms: Patch-Level Distillation Loss (PDL) and Patch-Level Feature Reconstruction method (PFR). Specifically, the PDL preserves feature-level consistency between successive models by implementing a distillation loss based on the contributions of patch tokens to new class learning. The PFR facilitates accurate classifier alignment by reconstructing old class features from previous tasks that adapt to new task knowledge. Extensive experiments validate the effectiveness of our DIA, revealing significant improvements on benchmark datasets in the NECIL setting, maintaining an optimal balance between computational complexity and accuracy. The full code implementation will be made publicly available upon the publication of this paper.
Deep learning has long been dominated by multi-layer perceptrons (MLPs), which have demonstrated superiority over other optimizable models in various domains. Recently, a new alternative to MLPs has emerged - Kolmogorov-Arnold Networks (KAN)- which are based on a fundamentally different mathematical framework. According to their authors, KANs address several major issues in MLPs, such as catastrophic forgetting in continual learning scenarios. However, this claim has only been supported by results from a regression task on a toy 1D dataset. In this paper, we extend the investigation by evaluating the performance of KANs in continual learning tasks within computer vision, specifically using the MNIST datasets. To this end, we conduct a structured analysis of the behavior of MLPs and two KAN-based models in a class-incremental learning scenario, ensuring that the architectures involved have the same number of trainable parameters. Our results demonstrate that an efficient version of KAN outperforms both traditional MLPs and the original KAN implementation. We further analyze the influence of hyperparameters in MLPs and KANs, as well as the impact of certain trainable parameters in KANs, such as bias and scale weights. Additionally, we provide a preliminary investigation of recent KAN-based convolutional networks and compare their performance with that of traditional convolutional neural networks. Our codes can be found at https://github.com/MrPio/KAN-Continual_Learning_tests.
We introduce a novel Recursive Fusion model, dubbed ReFu, designed to integrate point clouds and meshes for exemplar-free 3D Class-Incremental Learning, where the model learns new 3D classes while retaining knowledge of previously learned ones. Unlike existing methods that either rely on storing historical data to mitigate forgetting or focus on single data modalities, ReFu eliminates the need for exemplar storage while utilizing the complementary strengths of both point clouds and meshes. To achieve this, we introduce a recursive method which continuously accumulates knowledge by updating the regularized auto-correlation matrix. Furthermore, we propose a fusion module, featuring a Pointcloud-guided Mesh Attention Layer that learns correlations between the two modalities. This mechanism effectively integrates point cloud and mesh features, leading to more robust and stable continual learning. Experiments across various datasets demonstrate that our proposed framework outperforms existing methods in 3D class-incremental learning. Project Page: https://arlo397.github.io/ReFu/
Few-shot class-incremental learning (FSCIL) aims to incrementally recognize new classes using a few samples while maintaining the performance on previously learned classes. One of the effective methods to solve this challenge is to construct prototypical evolution classifiers. Despite the advancement achieved by most existing methods, the classifier weights are simply initialized using mean features. Because representations for new classes are weak and biased, we argue such a strategy is suboptimal. In this paper, we tackle this issue from two aspects. Firstly, thanks to the development of foundation models, we employ a foundation model, the CLIP, as the network pedestal to provide a general representation for each class. Secondly, to generate a more reliable and comprehensive instance representation, we propose a Knowledge Adapter (KA) module that summarizes the data-specific knowledge from training data and fuses it into the general representation. Additionally, to tune the knowledge learned from the base classes to the upcoming classes, we propose a mechanism of Incremental Pseudo Episode Learning (IPEL) by simulating the actual FSCIL. Taken together, our proposed method, dubbed as Knowledge Adaptation Network (KANet), achieves competitive performance on a wide range of datasets, including CIFAR100, CUB200, and ImageNet-R.