In class incremental learning (CIL) a model must learn new classes in a sequential manner without forgetting old ones. However, conventional CIL methods consider a balanced distribution for each new task, which ignores the prevalence of long-tailed distributions in the real world. In this work we propose two long-tailed CIL scenarios, which we term ordered and shuffled LT-CIL. Ordered LT-CIL considers the scenario where we learn from head classes collected with more samples than tail classes which have few. Shuffled LT-CIL, on the other hand, assumes a completely random long-tailed distribution for each task. We systematically evaluate existing methods in both LT-CIL scenarios and demonstrate very different behaviors compared to conventional CIL scenarios. Additionally, we propose a two-stage learning baseline with a learnable weight scaling layer for reducing the bias caused by long-tailed distribution in LT-CIL and which in turn also improves the performance of conventional CIL due to the limited exemplars. Our results demonstrate the superior performance (up to 6.44 points in average incremental accuracy) of our approach on CIFAR-100 and ImageNet-Subset. The code is available at https://github.com/xialeiliu/Long-Tailed-CIL
Class-incremental learning for semantic segmentation (CiSS) is presently a highly researched field which aims at updating a semantic segmentation model by sequentially learning new semantic classes. A major challenge in CiSS is overcoming the effects of catastrophic forgetting, which describes the sudden drop of accuracy on previously learned classes after the model is trained on a new set of classes. Despite latest advances in mitigating catastrophic forgetting, the underlying causes of forgetting specifically in CiSS are not well understood. Therefore, in a set of experiments and representational analyses, we demonstrate that the semantic shift of the background class and a bias towards new classes are the major causes of forgetting in CiSS. Furthermore, we show that both causes mostly manifest themselves in deeper classification layers of the network, while the early layers of the model are not affected. Finally, we demonstrate how both causes are effectively mitigated utilizing the information contained in the background, with the help of knowledge distillation and an unbiased cross-entropy loss.
Inspired by Regularized Lottery Ticket Hypothesis (RLTH), which hypothesizes that there exist smooth (non-binary) subnetworks within a dense network that achieve the competitive performance of the dense network, we propose a few-shot class incremental learning (FSCIL) method referred to as \emph{Soft-SubNetworks (SoftNet)}. Our objective is to learn a sequence of sessions incrementally, where each session only includes a few training instances per class while preserving the knowledge of the previously learned ones. SoftNet jointly learns the model weights and adaptive non-binary soft masks at a base training session in which each mask consists of the major and minor subnetwork; the former aims to minimize catastrophic forgetting during training, and the latter aims to avoid overfitting to a few samples in each new training session. We provide comprehensive empirical validations demonstrating that our SoftNet effectively tackles the few-shot incremental learning problem by surpassing the performance of state-of-the-art baselines over benchmark datasets.
Catastrophic forgetting has been a significant problem hindering the deployment of deep learning algorithms in the continual learning setting. Numerous methods have been proposed to address the catastrophic forgetting problem where an agent loses its generalization power of old tasks while learning new tasks. We put forward an alternative strategy to handle the catastrophic forgetting with knowledge amalgamation (CFA), which learns a student network from multiple heterogeneous teacher models specializing in previous tasks and can be applied to current offline methods. The knowledge amalgamation process is carried out in a single-head manner with only a selected number of memorized samples and no annotations. The teachers and students do not need to share the same network structure, allowing heterogeneous tasks to be adapted to a compact or sparse data representation. We compare our method with competitive baselines from different strategies, demonstrating our approach's advantages.
This paper studies class incremental learning (CIL) of continual learning (CL). Many approaches have been proposed to deal with catastrophic forgetting (CF) in CIL. Most methods incrementally construct a single classifier for all classes of all tasks in a single head network. To prevent CF, a popular approach is to memorize a small number of samples from previous tasks and replay them during training of the new task. However, this approach still suffers from serious CF as the parameters learned for previous tasks are updated or adjusted with only the limited number of saved samples in the memory. This paper proposes an entirely different approach that builds a separate classifier (head) for each task (called a multi-head model) using a transformer network, called MORE. Instead of using the saved samples in memory to update the network for previous tasks/classes in the existing approach, MORE leverages the saved samples to build a task specific classifier (adding a new classification head) without updating the network learned for previous tasks/classes. The model for the new task in MORE is trained to learn the classes of the task and also to detect samples that are not from the same data distribution (i.e., out-of-distribution (OOD)) of the task. This enables the classifier for the task to which the test instance belongs to produce a high score for the correct class and the classifiers of other tasks to produce low scores because the test instance is not from the data distributions of these classifiers. Experimental results show that MORE outperforms state-of-the-art baselines and is also naturally capable of performing OOD detection in the continual learning setting.
The human brain is capable of learning tasks sequentially mostly without forgetting. However, deep neural networks (DNNs) suffer from catastrophic forgetting when learning one task after another. We address this challenge considering a class-incremental learning scenario where the DNN sees test data without knowing the task from which this data originates. During training, Continual-Prune-and-Select (CP&S) finds a subnetwork within the DNN that is responsible for solving a given task. Then, during inference, CP&S selects the correct subnetwork to make predictions for that task. A new task is learned by training available neuronal connections of the DNN (previously untrained) to create a new subnetwork by pruning, which can include previously trained connections belonging to other subnetwork(s) because it does not update shared connections. This enables to eliminate catastrophic forgetting by creating specialized regions in the DNN that do not conflict with each other while still allowing knowledge transfer across them. The CP&S strategy is implemented with different subnetwork selection strategies, revealing superior performance to state-of-the-art continual learning methods tested on various datasets (CIFAR-100, CUB-200-2011, ImageNet-100 and ImageNet-1000). In particular, CP&S is capable of sequentially learning 10 tasks from ImageNet-1000 keeping an accuracy around 94% with negligible forgetting, a first-of-its-kind result in class-incremental learning. To the best of the authors' knowledge, this represents an improvement in accuracy above 10% when compared to the best alternative method.
In class-incremental learning, the model is expected to learn new classes continually while maintaining knowledge on previous classes. The challenge here lies in preserving the model's ability to effectively represent prior classes in the feature space, while adapting it to represent incoming new classes. We propose two distillation-based objectives for class incremental learning that leverage the structure of the feature space to maintain accuracy on previous classes, as well as enable learning the new classes. In our first objective, termed cross-space clustering (CSC), we propose to use the feature space structure of the previous model to characterize directions of optimization that maximally preserve the class: directions that all instances of a specific class should collectively optimize towards, and those that they should collectively optimize away from. Apart from minimizing forgetting, this indirectly encourages the model to cluster all instances of a class in the current feature space, and gives rise to a sense of herd-immunity, allowing all samples of a class to jointly combat the model from forgetting the class. Our second objective termed controlled transfer (CT) tackles incremental learning from an understudied perspective of inter-class transfer. CT explicitly approximates and conditions the current model on the semantic similarities between incrementally arriving classes and prior classes. This allows the model to learn classes in such a way that it maximizes positive forward transfer from similar prior classes, thus increasing plasticity, and minimizes negative backward transfer on dissimilar prior classes, whereby strengthening stability. We perform extensive experiments on two benchmark datasets, adding our method (CSCCT) on top of three prominent class-incremental learning methods. We observe consistent performance improvement on a variety of experimental settings.
In class incremental learning (CIL) setting, groups of classes are introduced to a model in each learning phase. The goal is to learn a unified model performant on all the classes observed so far. Given the recent popularity of Vision Transformers (ViTs) in conventional classification settings, an interesting question is to study their continual learning behaviour. In this work, we develop a Debiased Dual Distilled Transformer for CIL dubbed $\textrm{D}^3\textrm{Former}$. The proposed model leverages a hybrid nested ViT design to ensure data efficiency and scalability to small as well as large datasets. In contrast to a recent ViT based CIL approach, our $\textrm{D}^3\textrm{Former}$ does not dynamically expand its architecture when new tasks are learned and remains suitable for a large number of incremental tasks. The improved CIL behaviour of $\textrm{D}^3\textrm{Former}$ owes to two fundamental changes to the ViT design. First, we treat the incremental learning as a long-tail classification problem where the majority samples from new classes vastly outnumber the limited exemplars available for old classes. To avoid the bias against the minority old classes, we propose to dynamically adjust logits to emphasize on retaining the representations relevant to old tasks. Second, we propose to preserve the configuration of spatial attention maps as the learning progresses across tasks. This helps in reducing catastrophic forgetting by constraining the model to retain the attention on the most discriminative regions. $\textrm{D}^3\textrm{Former}$ obtains favorable results on incremental versions of CIFAR-100, MNIST, SVHN, and ImageNet datasets. Code is available at https://tinyurl.com/d3former
The continual appearance of new objects in the visual world poses considerable challenges for current deep learning methods in real-world deployments. The challenge of new task learning is often exacerbated by the scarcity of data for the new categories due to rarity or cost. Here we explore the important task of Few-Shot Class-Incremental Learning (FSCIL) and its extreme data scarcity condition of one-shot. An ideal FSCIL model needs to perform well on all classes, regardless of their presentation order or paucity of data. It also needs to be robust to open-set real-world conditions and be easily adapted to the new tasks that always arise in the field. In this paper, we first reevaluate the current task setting and propose a more comprehensive and practical setting for the FSCIL task. Then, inspired by the similarity of the goals for FSCIL and modern face recognition systems, we propose our method -- Augmented Angular Loss Incremental Classification or ALICE. In ALICE, instead of the commonly used cross-entropy loss, we propose to use the angular penalty loss to obtain well-clustered features. As the obtained features not only need to be compactly clustered but also diverse enough to maintain generalization for future incremental classes, we further discuss how class augmentation, data augmentation, and data balancing affect classification performance. Experiments on benchmark datasets, including CIFAR100, miniImageNet, and CUB200, demonstrate the improved performance of ALICE over the state-of-the-art FSCIL methods.
Few-shot class-incremental learning (FSCIL) has been proposed aiming to enable a deep learning system to incrementally learn new classes with limited data. Recently, a pioneer claims that the commonly used replay-based method in class-incremental learning (CIL) is ineffective and thus not preferred for FSCIL. This has, if truth, a significant influence on the fields of FSCIL. In this paper, we show through empirical results that adopting the data replay is surprisingly favorable. However, storing and replaying old data can lead to a privacy concern. To address this issue, we alternatively propose using data-free replay that can synthesize data by a generator without accessing real data. In observing the the effectiveness of uncertain data for knowledge distillation, we impose entropy regularization in the generator training to encourage more uncertain examples. Moreover, we propose to relabel the generated data with one-hot-like labels. This modification allows the network to learn by solely minimizing the cross-entropy loss, which mitigates the problem of balancing different objectives in the conventional knowledge distillation approach. Finally, we show extensive experimental results and analysis on CIFAR-100, miniImageNet and CUB-200 to demonstrate the effectiveness of our proposed one.
Deep neural networks perform well on classification tasks where data streams are i.i.d. and labeled data is abundant. Challenges emerge with non-stationary training data streams such as continual learning. One powerful approach that has addressed this challenge involves pre-training of large encoders on volumes of readily available data, followed by task-specific tuning. Given a new task, however, updating the weights of these encoders is challenging as a large number of weights needs to be fine-tuned, and as a result, they forget information about the previous tasks. In the present work, we propose a model architecture to address this issue, building upon a discrete bottleneck containing pairs of separate and learnable key-value codes. Our paradigm will be to encode; process the representation via a discrete bottleneck; and decode. Here, the input is fed to the pre-trained encoder, the output of the encoder is used to select the nearest keys, and the corresponding values are fed to the decoder to solve the current task. The model can only fetch and re-use a sparse number of these key-value pairs during inference, enabling localized and context-dependent model updates. We theoretically investigate the ability of the discrete key-value bottleneck to minimize the effect of learning under distribution shifts and show that it reduces the complexity of the hypothesis class. We empirically verify the proposed method under challenging class-incremental learning scenarios and show that the proposed model - without any task boundaries - reduces catastrophic forgetting across a wide variety of pre-trained models, outperforming relevant baselines on this task.
Standard deep learning-based classification approaches require collecting all samples from all classes in advance and are trained offline. This paradigm may not be practical in real-world clinical applications, where new classes are incrementally introduced through the addition of new data. Class incremental learning is a strategy allowing learning from such data. However, a major challenge is catastrophic forgetting, i.e., performance degradation on previous classes when adapting a trained model to new data. Prior methodologies to alleviate this challenge save a portion of training data require perpetual storage of such data that may introduce privacy issues. Here, we propose a novel data-free class incremental learning framework that first synthesizes data from the model trained on previous classes to generate a \ours. Subsequently, it updates the model by combining the synthesized data with new class data. Furthermore, we incorporate a cosine normalized Cross-entropy loss to mitigate the adverse effects of the imbalance, a margin loss to increase separation among previous classes and new ones, and an intra-domain contrastive loss to generalize the model trained on the synthesized data to real data. We compare our proposed framework with state-of-the-art methods in class incremental learning, where we demonstrate improvement in accuracy for the classification of 11,062 echocardiography cine series of patients.
Deep learning models suffer from catastrophic forgetting when learning new tasks incrementally. Incremental learning has been proposed to retain the knowledge of old classes while learning to identify new classes. A typical approach is to use a few exemplars to avoid forgetting old knowledge. In such a scenario, data imbalance between old and new classes is a key issue that leads to performance degradation of the model. Several strategies have been designed to rectify the bias towards the new classes due to data imbalance. However, they heavily rely on the assumptions of the bias relation between old and new classes. Therefore, they are not suitable for complex real-world applications. In this study, we propose an assumption-agnostic method, Multi-Granularity Regularized re-Balancing (MGRB), to address this problem. Re-balancing methods are used to alleviate the influence of data imbalance; however, we empirically discover that they would under-fit new classes. To this end, we further design a novel multi-granularity regularization term that enables the model to consider the correlations of classes in addition to re-balancing the data. A class hierarchy is first constructed by grouping the semantically or visually similar classes. The multi-granularity regularization then transforms the one-hot label vector into a continuous label distribution, which reflects the relations between the target class and other classes based on the constructed class hierarchy. Thus, the model can learn the inter-class relational information, which helps enhance the learning of both old and new classes. Experimental results on both public datasets and a real-world fault diagnosis dataset verify the effectiveness of the proposed method.
While deep neural networks (DNNs) have achieved impressive classification performance in closed-world learning scenarios, they typically fail to generalize to unseen categories in dynamic open-world environments, in which the number of concepts is unbounded. In contrast, human and animal learners have the ability to incrementally update their knowledge by recognizing and adapting to novel observations. In particular, humans characterize concepts via exclusive (unique) sets of essential features, which are used for both recognizing known classes and identifying novelty. Inspired by natural learners, we introduce a Sparse High-level-Exclusive, Low-level-Shared feature representation (SHELS) that simultaneously encourages learning exclusive sets of high-level features and essential, shared low-level features. The exclusivity of the high-level features enables the DNN to automatically detect out-of-distribution (OOD) data, while the efficient use of capacity via sparse low-level features permits accommodating new knowledge. The resulting approach uses OOD detection to perform class-incremental continual learning without known class boundaries. We show that using SHELS for novelty detection results in statistically significant improvements over state-of-the-art OOD detection approaches over a variety of benchmark datasets. Further, we demonstrate that the SHELS model mitigates catastrophic forgetting in a class-incremental learning setting,enabling a combined novelty detection and accommodation framework that supports learning in open-world settings
Few-shot class-incremental learning (FSCIL) aims to incrementally fine-tune a model (trained on base classes) for a novel set of classes using a few examples without forgetting the previous training. Recent efforts address this problem primarily on 2D images. However, due to the advancement of camera technology, 3D point cloud data has become more available than ever, which warrants considering FSCIL on 3D data. This paper addresses FSCIL in the 3D domain. In addition to well-known issues of catastrophic forgetting of past knowledge and overfitting of few-shot data, 3D FSCIL can bring newer challenges. For example, base classes may contain many synthetic instances in a realistic scenario. In contrast, only a few real-scanned samples (from RGBD sensors) of novel classes are available in incremental steps. Due to the data variation from synthetic to real, FSCIL endures additional challenges, degrading performance in later incremental steps. We attempt to solve this problem using Microshapes (orthogonal basis vectors) by describing any 3D objects using a pre-defined set of rules. It supports incremental training with few-shot examples minimizing synthetic to real data variation. We propose new test protocols for 3D FSCIL using popular synthetic datasets (ModelNet and ShapeNet) and 3D real-scanned datasets (ScanObjectNN and CO3D). By comparing state-of-the-art methods, we establish the effectiveness of our approach in the 3D domain.
Class-incremental learning (CIL) learns a classification model with training data of different classes arising progressively. Existing CIL either suffers from serious accuracy loss due to catastrophic forgetting, or invades data privacy by revisiting used exemplars. Inspired by linear learning formulations, we propose an analytic class-incremental learning (ACIL) with absolute memorization of past knowledge while avoiding breaching of data privacy (i.e., without storing historical data). The absolute memorization is demonstrated in the sense that class-incremental learning using ACIL given present data would give identical results to that from its joint-learning counterpart which consumes both present and historical samples. This equality is theoretically validated. Data privacy is ensured since no historical data are involved during the learning process. Empirical validations demonstrate ACIL's competitive accuracy performance with near-identical results for various incremental task settings (e.g., 5-50 phases). This also allows ACIL to outperform the state-of-the-art methods for large-phase scenarios (e.g., 25 and 50 phases).
With the tremendous expansion of graphs data, node classification shows its great importance in many real-world applications. Existing graph neural network based methods mainly focus on classifying unlabeled nodes within fixed classes with abundant labeling. However, in many practical scenarios, graph evolves with emergence of new nodes and edges. Novel classes appear incrementally along with few labeling due to its newly emergence or lack of exploration. In this paper, we focus on this challenging but practical graph few-shot class-incremental learning (GFSCIL) problem and propose a novel method called Geometer. Instead of replacing and retraining the fully connected neural network classifer, Geometer predicts the label of a node by finding the nearest class prototype. Prototype is a vector representing a class in the metric space. With the pop-up of novel classes, Geometer learns and adjusts the attention-based prototypes by observing the geometric proximity, uniformity and separability. Teacher-student knowledge distillation and biased sampling are further introduced to mitigate catastrophic forgetting and unbalanced labeling problem respectively. Experimental results on four public datasets demonstrate that Geometer achieves a substantial improvement of 9.46% to 27.60% over state-of-the-art methods.
Real-world applications require the classification model to adapt to new classes without forgetting old ones. Correspondingly, Class-Incremental Learning (CIL) aims to train a model with limited memory size to meet this requirement. Typical CIL methods tend to save representative exemplars from former classes to resist forgetting, while recent works find that storing models from history can substantially boost the performance. However, the stored models are not counted into the memory budget, which implicitly results in unfair comparisons. We find that when counting the model size into the total budget and comparing methods with aligned memory size, saving models do not consistently work, especially for the case with limited memory budgets. As a result, we need to holistically evaluate different CIL methods at different memory scales and simultaneously consider accuracy and memory size for measurement. On the other hand, we dive deeply into the construction of the memory buffer for memory efficiency. By analyzing the effect of different layers in the network, we find that shallow and deep layers have different characteristics in CIL. Motivated by this, we propose a simple yet effective baseline, denoted as MEMO for Memory-efficient Expandable MOdel. MEMO extends specialized layers based on the shared generalized representations, efficiently extracting diverse representations with modest cost and maintaining representative exemplars. Extensive experiments on benchmark datasets validate MEMO's competitive performance. Code is available at: https://github.com/wangkiw/ICLR23-MEMO
Deep learning models have shown a great effectiveness in recognition of findings in medical images. However, they cannot handle the ever-changing clinical environment, bringing newly annotated medical data from different sources. To exploit the incoming streams of data, these models would benefit largely from sequentially learning from new samples, without forgetting the previously obtained knowledge. In this paper we introduce LifeLonger, a benchmark for continual disease classification on the MedMNIST collection, by applying existing state-of-the-art continual learning methods. In particular, we consider three continual learning scenarios, namely, task and class incremental learning and the newly defined cross-domain incremental learning. Task and class incremental learning of diseases address the issue of classifying new samples without re-training the models from scratch, while cross-domain incremental learning addresses the issue of dealing with datasets originating from different institutions while retaining the previously obtained knowledge. We perform a thorough analysis of the performance and examine how the well-known challenges of continual learning, such as the catastrophic forgetting exhibit themselves in this setting. The encouraging results demonstrate that continual learning has a major potential to advance disease classification and to produce a more robust and efficient learning framework for clinical settings. The code repository, data partitions and baseline results for the complete benchmark will be made publicly available.
The ability to learn new concepts continually is necessary in this ever-changing world. However, deep neural networks suffer from catastrophic forgetting when learning new categories. Many works have been proposed to alleviate this phenomenon, whereas most of them either fall into the stability-plasticity dilemma or take too much computation or storage overhead. Inspired by the gradient boosting algorithm to gradually fit the residuals between the target model and the previous ensemble model, we propose a novel two-stage learning paradigm FOSTER, empowering the model to learn new categories adaptively. Specifically, we first dynamically expand new modules to fit the residuals between the target and the output of the original model. Next, we remove redundant parameters and feature dimensions through an effective distillation strategy to maintain the single backbone model. We validate our method FOSTER on CIFAR-100 and ImageNet-100/1000 under different settings. Experimental results show that our method achieves state-of-the-art performance. Code is available at: https://github.com/G-U-N/ECCV22-FOSTER.