Catastrophic Forgetting Research Articles

CCSI: Continual Class-Specific Impression for data-free class incremental learning

In real-world clinical settings, traditional deep learning-based classification methods struggle with diagnosing newly introduced disease types because they require samples from all disease classes for offline training. Class incremental learning offers a promising solution by adapting a deep network trained on specific disease classes to handle new diseases. However, catastrophic forgetting occurs, decreasing the performance of earlier classes when adapting the model to new data. Prior proposed methodologies to overcome this require perpetual storage of previous samples, posing potential practical concerns regarding privacy and storage regulations in healthcare. To this end, we propose a novel data-free class incremental learning framework that utilizes data synthesis on learned classes instead of data storage from previous classes. Our key contributions include acquiring synthetic data known as Continual Class-Specific Impression (CCSI) for previously inaccessible trained classes and presenting a methodology to effectively utilize this data for updating networks when introducing new classes. We obtain CCSI by employing data inversion over gradients of the trained classification model on previous classes starting from the mean image of each class inspired by common landmarks shared among medical images and utilizing continual normalization layers statistics as a regularizer in this pixel-wise optimization process. Subsequently, we update the network by combining the synthesized data with new class data and incorporate several losses, including an intra-domain contrastive loss to generalize the deep network trained on the synthesized data to real data, a margin loss to increase separation among previous classes and new ones, and a cosine-normalized cross-entropy loss to alleviate the adverse effects of imbalanced distributions in training data. Extensive experiments show that the proposed framework achieves state-of-the-art performance on four of the public MedMNIST datasets and in-house echocardiography cine series, with an improvement in classification accuracy of up to 51% compared to baseline data-free methods. Our code is available at https://github.com/ubc-tea/Continual-Impression-CCSI.

RRA-FFSCIL: Inter-intra classes representation and relationship augmentation federated few-shot incremental learning

Federated learning (FL), as a distributed machine learning paradigm, enables on-device model training and inference without data updates or privacy breaches, promoting edge intelligence in Industrial Internet of Things (IIoTs) applications. However, FL applying in IIoTs faces some challenges in practice: (1) the catastrophic forgetting of old data resulting from a streaming reception for arriving data; (2) the forgetting exacerbated by the non-i.i.d. data distribution across devices; (3) the overfitting caused by insufficient sample size of new data (i.e., few samples). To address the above problems, we propose an inter-intra classes representation and relationship augmentation federated few-shot incremental learning framework called RRA-FFSCIL, effectively mitigating catastrophic forgetting from local and global perspectives. Specifically, forgetting from streaming data is alleviated by designed modules in the client side of RRA-FFSCIL: a random few-shot (FS) episodic adaptation strategy, a class-projection relation strengthening loss function, and a class-aware distribution balancing loss function. Meanwhile, the mitigation of global forgetting from non-i.i.d. of distributions across devices thanks to selecting the best old models on the server to assist local class relation enhancement. We conduct extensive experiments on three datasets (cifar100, tinyImageNet, miniImageNet) to validate the effectiveness and superiority of RRA-FFSCIL in addressing the forgetting problem associated with streaming data.

Continuous unsupervised domain adaptation using stabilized representations and experience replay

We introduce an algorithm for tackling the problem of unsupervised domain adaptation (UDA) in continual learning (CL) scenarios. The primary objective is to maintain model generalization under domain shift when new domains with solely unannotated data arrive continually. Under this setting, While there are many existing UDA algorithms, we have access to only the current target domain dataset simultaneously and need to update the model to learn a task with unannotated data without forgetting the tasks learnt before. Our contribution is to address this learning setting which faces challenges seen in both UDA and CL. The core assumption in the learning setting is that the tasks share the same classes which enables benefiting from knowledge transfer. Our solution is based on stabilizing the learned internal distribution after learning the initial task to enhances the model generalization on new domains. The internal distribution is modeled by network responses in the eventual model hidden layer. We model this internal distribution using a Gaussian mixture model (GMM ) and update the model by matching the internally learned distribution of new domains to the estimated GMM as a surrogate for the past learned internal distribution. Additionally, we leverage experience replay to overcome the problem of catastrophic forgetting, where the model loses previously acquired knowledge when learning new tasks. We offer theoretical analysis to explain why our algorithm would work. We also offer comparative and analytic experiments to demonstrate that our method is effective. Our implementation code is available at: https://github.com/rostami-m/LDACID.

Cross-domain continual learning via CLAMP

Artificial neural networks, celebrated for their human-like cognitive learning abilities, often encounter the well-known catastrophic forgetting (CF) problem, where the neural networks lose the proficiency in previously acquired knowledge. Despite numerous efforts to mitigate CF, it remains the significant challenge particularly in complex changing environments. This challenge is even more pronounced in cross-domain adaptation following the continual learning (CL) setting, which is a more challenging and realistic scenario that is under-explored. To this end, this article proposes a cross-domain CL approach making possible to deploy a single model in such environments without additional labelling costs. Our approach, namely continual learning approach for many processes (CLAMP), integrates a class-aware adversarial domain adaptation strategy to align a source domain and a target domain. An assessor-guided learning process is put forward to navigate the learning process of a base model assigning a set of weights to every sample controlling the influence of every sample and the interactions of each loss function in such a way to balance the stability and plasticity dilemma thus preventing the CF problem. The first assessor focuses on the negative transfer problem rejecting irrelevant samples of the source domain while the second assessor prevents noisy pseudo labels of the target domain. Both assessors are trained in the meta-learning approach using random transformation techniques and similar samples of the source domain. Theoretical analysis and extensive numerical validations demonstrate that CLAMP significantly outperforms established baseline algorithms across all experiments by at least 10% margin.

Two-dimensional hybrid incremental learning (2DHIL) framework for semantic segmentation of skin tissues

This study aims to enhance the robustness and generalization capability of a deep learning transformer model used for segmenting skin carcinomas and tissues through the introduction of incremental learning. Deep learning AI models demonstrate their claimed performance only for tasks and data types for which they are specifically trained. Their performance is severely challenged for the test cases which are not similar to training data thus questioning their robustness and ability to generalize. Moreover, these models require an enormous amount of annotated data for training to achieve desired performance. The availability of large annotated data, particularly for medical applications, is itself a challenge. Despite efforts to alleviate this limitation through techniques like data augmentation, transfer learning, and few-shot training, the challenge persists. To address this, we propose refining the models incrementally as new classes are discovered and more data becomes available, emulating the human learning process. However, deep learning models face the challenge of catastrophic forgetting during incremental training. Therefore, we introduce a two-dimensional hybrid incremental learning framework for segmenting non-melanoma skin cancers and tissues from histopathology images. Our approach involves progressively adding new classes and introducing data of varying specifications to introduce adaptability in the models. We also employ a combination of loss functions to facilitate new learning and mitigate catastrophic forgetting. Our extended experiments demonstrate significant improvements, with an F1 score reaching 91.78, mIoU of 93.00, and an average accuracy of 95%. These findings highlight the effectiveness of our incremental learning strategy in enhancing the robustness and generalization of deep learning segmentation models while mitigating catastrophic forgetting.

CIU-L: A class-incremental learning and machine unlearning passive sensing system for human identification

With the development of passive sensing technology, WiFi-based identification research has attracted much attention in areas such as human–computer interaction and home security. Although WiFi sensing-based human identification has achieved initial success, it is currently mainly applicable to scenarios where the user’s identity category is fixed and not applicable to scenarios where the user’s identity category changes frequently. In this paper, we propose an identification system (CIU-L) in a scenario where user’s identity categories frequently change, allowing for incremental registration and unregistration of identity categories. To the best of our knowledge, this is the first attempt to register and unregister user identity information under the previous identity category constraints. CIU-L proposes a training and updating strategy in the registration phase of new user to avoid catastrophic forgetting of old user’s identity information, and trains a targeted noise for the user to be unregistered in the unregistration phase of old user, achieving precise removal of the user to be unregistered without affecting the retained users. In addition, this paper presents adequate comparative experiments of CIU-L with other systems in the user identity category fixing scenario. The experimental results show that the average difference between CIU-L and other systems in terms of Accuracy, Precision, Recall and F1-Score is within 5% of each other, while running time and storage space are saved by more than 6 times, which is more capable of meeting the needs of identity recognition in real scenarios.

Few-shot class incremental learning via robust transformer approach

Few-Shot Class-Incremental Learning (FSCIL)presents an extension of the Class Incremental Learning (CIL)problem where a model is faced with the problem of data scarcity while addressing the Catastrophic Forgetting (CF)problem. This problem remains an open problem because all recent works are built upon the Convolutional Neural Networks (CNNs)performing sub-optimally compared to the transformer approaches. Our paper presents Robust Transformer Approach (ROBUSTA)built upon the Compact Convolutional Transformer (CCT). 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 CFis 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 ROBUSTAis 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.

Adaptive knowledge transfer for class incremental learning

Humans are excellent at adapting to constantly changing circumstances, but deep neural networks have catastrophic forgetting. Recently, significant progress has been made with class-incremental methods based on dynamic network structures. However, these methods combine individual networks in a simplistic coupled manner, neglecting the fusion capabilities between modules, leading to a decline in the overall prediction performance. To address this, we propose a class-incremental learning paradigm with adaptive knowledge transfer. This paradigm leverages crucial self-learning factors to transfer the importance knowledge of old classes as much as possible, allowing each module to integrate the optimal information from the current class. Experiments demonstrate that our designed adaptive knowledge transfer module effectively reduces the sharpness of decision boundaries, thereby significantly improving the final accuracy. Additionally, we have devised a compression module with supplementary learning to mitigate errors arising from long-term session sequences during model fusion. Extensive experiments conducted on benchmarks, Our approach achieved an average accuracy exceeding 2.72%, 1.30% in the CIFAR-100 and ImageNet-100/1000 achieves SOTA performance in both ordinary and challenging class-incremental settings.

A memory-friendly class-incremental learning method for hand gesture recognition using HD-sEMG

Hand gesture recognition (HGR) plays a vital role in human-computer interaction. The integration of high-density surface electromyography (HD-sEMG) and deep neural networks (DNNs) has significantly improved the robustness and accuracy of HGR systems. These methods are typically effective for a fixed set of trained gestures. However, the need for new gesture classes over time poses a challenge. Introducing new classes to DNNs can lead to a substantial decrease in accuracy for previously learned tasks, a phenomenon known as “catastrophic forgetting,” especially when the training data for earlier tasks is not retained and retrained. This issue is exacerbated in embedded devices with limited storage, which struggle to store the large-scale data of HD-sEMG. Class-incremental learning (CIL) is an effective method to reduce catastrophic forgetting. However, existing CIL methods for HGR rarely focus on reducing memory load. To address this, we propose a memory-friendly CIL method for HGR using HD-sEMG. Our approach includes a lightweight convolutional neural network, named SeparaNet, for feature representation learning, coupled with a nearest-mean-of-exemplars classifier for classification. We introduce a priority exemplar selection algorithm inspired by the herding effect to maintain a manageable set of exemplars during training. Furthermore, a task-equal-weight exemplar sampling strategy is proposed to effectively reduce memory load while preserving high recognition performance. Experimental results on two datasets demonstrate that our method significantly reduces the number of retained exemplars to only a quarter of that required by other CIL methods, accounting for less than 5 ​% of the total samples, while still achieving comparable average accuracy.

Human-like Learning in Temporally Structured Environments

Natural environments have correlations at a wide range of timescales. Human cognition is tuned to this temporal structure, as seen by power laws of learning and memory, and by spacing effects whereby the intervals between repeated training data affect how long knowledge is retained. Machine learning is instead dominated by batch iid training or else relatively simple nonstationarity assumptions such as random walks or discrete task sequences. The main contributions of our work are: (1) We develop a Bayesian model formalizing the brain's inductive bias for temporal structure and show our model accounts for key features of human learning and memory. (2) We translate the model into a new gradient-based optimization technique for neural networks that endows them with human-like temporal inductive bias and improves their performance in realistic nonstationary tasks. Our technical approach is founded on Bayesian inference over 1/f noise, a statistical signature of many natural environments with long-range, power law correlations. We derive a new closed-form solution to this problem by treating the state of the environment as a sum of processes on different timescales and applying an extended Kalman filter to learn all timescales jointly. We then derive a variational approximation of this model for training neural networks, which can be used as a drop-in replacement for standard optimizers in arbitrary architectures. Our optimizer decomposes each weight in the network as a sum of subweights with different learning and decay rates and tracks their joint uncertainty. Thus knowledge becomes distributed across timescales, enabling rapid adaptation to task changes while retaining long-term knowledge and avoiding catastrophic interference. Simulations show improved performance in environments with realistic multiscale nonstationarity. Finally, we present simulations showing our model gives essentially parameter-free fits of learning, forgetting, and spacing effects in human data. We then explore the analogue of human spacing effects in a deep net trained in a structured environment where tasks recur at different rates and compare the model's behavioral properties to those of people.

A multifidelity approach to continual learning for physical systems

We introduce a novel continual learning method based on multifidelity deep neural networks. This method learns the correlation between the output of previously trained models and the desired output of the model on the current training dataset, limiting catastrophic forgetting. On its own the multifidelity continual learning method shows robust results that limit forgetting across several datasets. Additionally, we show that the multifidelity method can be combined with existing continual learning methods, including replay and memory aware synapses, to further limit catastrophic forgetting. The proposed continual learning method is especially suited for physical problems where the data satisfy the same physical laws on each domain, or for physics-informed neural networks, because in these cases we expect there to be a strong correlation between the output of the previous model and the model on the current training domain.

Analysis of Continual Learning Techniques for Image Generative Models with Learned Class Information Management.

The advancements in deep learning have significantly enhanced the capability of image generation models to produce images aligned with human intentions. However, training and adapting these models to new data and tasks remain challenging because of their complexity and the risk of catastrophic forgetting. This study proposes a method for addressing these challenges involving the application of class-replacement techniques within a continual learning framework. This method utilizes selective amnesia (SA) to efficiently replace existing classes with new ones while retaining crucial information. This approach improves the model's adaptability to evolving data environments while preventing the loss of past information. We conducted a detailed evaluation of class-replacement techniques, examining their impact on the "class incremental learning" performance of models and exploring their applicability in various scenarios. The experimental results demonstrated that our proposed method could enhance the learning efficiency and long-term performance of image generation models. This study broadens the application scope of image generation technology and supports the continual improvement and adaptability of corresponding models.

Bridging pre-trained models to continual learning: A hypernetwork based framework with parameter-efficient fine-tuning techniques

Modern techniques of pre-training and fine-tuning have significantly improved the performance of models on downstream tasks. However, this improvement faces challenges when pre-trained models encounter the necessity to adapt sequentially to multiple downstream tasks within the context of continuously shifting training data. In this study, we aim to leverage the general capabilities of pre-trained models for knowledge sharing across different tasks while endow them with the capability for continuous learning. To this end, we propose a Hypernetwork-based Parameter Efficient Fine-Tuning (HyperPEFT) framework. Utilizing a pre-trained Vision Transformer (ViT) as the backbone, HyperPEFT is capable of incorporating various PEFT techniques, enabling the pre-trained ViT to adapt to diverse downstream tasks. The core of our method lies in the application of hypernetworks, which efficiently encapsulate task-specific information, significantly reducing task interference and fortifying the model against catastrophic forgetting. The adoption PEFT techniques allows for precise adjustments to the pre-trained models, enhancing their performance for each specific task. Moreover, this strategy employs a shared hypernetwork to make task-specific adjustments, thereby facilitating knowledge sharing across different tasks for pre-trained models. The extensive experiments reveal that our method effectively mitigates catastrophic forgetting, outperforms comparison methods, and uncovers latent associations among tasks. Overall, this study introduces a unified strategy that synergistically blends the general capabilities of pre-trained models with the necessary adaptability for continual learning scenarios.

TLCE: Transfer-Learning Based Classifier Ensembles for Few-Shot Class-Incremental Learning

Few-shot class-incremental learning (FSCIL) struggles to incrementally recognize novel classes from few examples without catastrophic forgetting of old classes or overfitting to new classes. We propose TLCE, which ensembles multiple pre-trained models to improve separation of novel and old classes. Specifically, we use episodic training to map images from old classes to quasi-orthogonal prototypes, which minimizes interference between old and new classes. Then, we incorporate the use of ensembling diverse pre-trained models to further tackle the challenge of data imbalance and enhance adaptation to novel classes. Extensive experiments on various datasets demonstrate that our transfer learning ensemble approach outperforms state-of-the-art FSCIL methods.

Semi-supervised few-shot class-incremental learning based on dynamic topology evolution

Incremental learning of new classes is crucial to developing real-world artificial intelligence systems. However, in some cases, we can only use a limited number of images to train the incremental model due to the difficulty of obtaining them. Few-shot class-incremental learning (FSCIL) is a new paradigm of incremental learning, which continually learns new concepts from a few examples. This setup can quickly lead to catastrophic forgetting and overfitting problems, severely affecting model performance. To address these problems, we propose a dynamic topological evolution (DTE) model that unifies knowledge of old and new categories. In this work, we focus on a novel semi-supervised few-shot class-incremental learning (SSFSCIL) to incrementally learn new categories with a limited number of labeled samples and a large number of unlabeled samples. In detail, we use the Voronoi polygon criterion to divide the feature space and construct a Hebbian graph to represent topological relationships between classes. Then, the corresponding k-nearest neighbor subgraph is constructed for the new incremental task, and label propagation of graph nodes is carried out on the unlabeled set to generate the corresponding pseudo labels. Finally, the new category knowledge is evolved into the actual category knowledge by integrating the Hebbian graph and the k-nearest neighbor subgraph obtained in the new incremental task. Extensive experiments on the mini-ImageNet and CIFAR100 datasets show that the accuracy of the DTE outperforms the baseline and improves by 1.41% and 1.13% on the last incremental task. Furthermore, our methodology achieved a 35.6% time improvement compared to full-supervised methods on the CIFAR100 dataset. In addition to the experiments on the two datasets mentioned above, the validation of our proposed DTE model extends to the engineering application domain of the RESICS-45 aircraft image scene classification dataset.

Learning At a Glance: Towards Interpretable Data-Limited Continual Semantic Segmentation Via Semantic-Invariance Modelling.

Continual semantic segmentation (CSS) based on incremental learning (IL) is a great endeavour in developing human- like segmentation models. However, current CSS approaches encounter challenges in the trade-off between preserving old knowledge and learning new ones, where they still need large-scale annotated data for incremental training and lack interpretability. In this paper, we present Learning at a Glance (LAG), an efficient, robust, human- like and interpretable approach for CSS. Specifically, LAG is a simple and model-agnostic architecture, yet it achieves competitive CSS efficiency with limited incremental data. Inspired by human- like recognition patterns, we propose a semantic-invariance modelling approach via semantic features decoupling that simultaneously reconciles solid knowledge inheritance and new-term learning. Concretely, the proposed decoupling manner includes two ways, i.e., channel- wise decoupling and spatial-level neuron-relevant semantic consistency. Our approach preserves semantic-invariant knowledge as solid prototypes to alleviate catastrophic forgetting, while also constraining sample-specific contents through an asymmetric contrastive learning method to enhance model robustness during IL steps. Experimental results in multiple datasets validate the effectiveness of the proposed method. Furthermore, we introduce a novel CSS protocol that better reflects realistic data-limited CSS settings, and LAG achieves superior performance under multiple data-limited conditions.

FedSKF: Selective Knowledge Fusion via Optimal Transport in Federated Class Incremental Learning

Federated learning has been a hot topic in the field of artificial intelligence in recent years due to its distributed nature and emphasis on privacy protection. To better align with real-world scenarios, federated class incremental learning (FCIL) has emerged as a new research trend, but it faces challenges such as heterogeneous data, catastrophic forgetting, and inter-client interference. However, most existing methods enhance model performance at the expense of privacy, such as uploading prototypes or samples, which violates the basic principle of only transmitting models in federated learning. This paper presents a novel selective knowledge fusion (FedSKF) model to address data heterogeneity and inter-client interference without sacrificing any privacy. Specifically, this paper introduces a PIT (projection in turn) module on the server side to indirectly recover client data distribution information through optimal transport. Subsequently, to reduce inter-client interference, knowledge of the global model is selectively absorbed via knowledge distillation and an incomplete synchronization classifier at the client side, namely an SKS (selective knowledge synchronization) module. Furthermore, to mitigate global catastrophic forgetting, a global forgetting loss is proposed to distill knowledge from the old global model. Our framework can easily integrate various CIL methods, allowing it to adapt to application scenarios with varying privacy requirements. We conducted extensive experiments on CIFAR100 and Tiny-ImageNet datasets, and the performance of our method surpasses existing works.

Learning vision-based robotic manipulation tasks sequentially in offline reinforcement learning settings

Abstract With the rise of deep reinforcement learning (RL) methods, many complex robotic manipulation tasks are being solved. However, harnessing the full power of deep learning requires large datasets. Online RL does not suit itself readily into this paradigm due to costly and time-consuming agent-environment interaction. Therefore, many offline RL algorithms have recently been proposed to learn robotic tasks. But mainly, all such methods focus on a single-task or multitask learning, which requires retraining whenever we need to learn a new task. Continuously learning tasks without forgetting previous knowledge combined with the power of offline deep RL would allow us to scale the number of tasks by adding them one after another. This paper investigates the effectiveness of regularisation-based methods like synaptic intelligence for sequentially learning image-based robotic manipulation tasks in an offline-RL setup. We evaluate the performance of this combined framework against common challenges of sequential learning: catastrophic forgetting and forward knowledge transfer. We performed experiments with different task combinations to analyse the effect of task ordering. We also investigated the effect of the number of object configurations and the density of robot trajectories. We found that learning tasks sequentially helps in the retention of knowledge from previous tasks, thereby reducing the time required to learn a new task. Regularisation-based approaches for continuous learning, like the synaptic intelligence method, help mitigate catastrophic forgetting but have shown only limited transfer of knowledge from previous tasks.

CoLLaRS : A cloud–edge–terminal collaborative lifelong learning framework for AIoT

AIoT applications often encounter challenges such as terminal resource constraints, data drift, and data heterogeneity in real world, leading to problems such as catastrophic forgetting, low generalization ability, and low accuracy during model training. To address these challenges, we proposed CoLLaRS, a cloud–edge–terminal collaborative lifelong learning framework for AIoT applications. In the CoLLaRS framework, we alleviate the problem of terminal resource constraints by uploading terminal tasks at the edge. CoLLaRS uses continuous training at the edge to achieve lifelong learning training of the model and solve the problem of catastrophic forgetting. CoLLaRS employs federated optimization in the cloud to perform personalized aggregation of different edge models and solve the problem of weak model generalization ability. Finally, the model is fine-tuned at the terminal to further optimize its accuracy in local tasks. Our experiments on real-world datasets showed that CoLLaRS has an 8% improvement in accuracy and a 5% improvement in backward transfer(BWT) and forward transfer(FWT) compared to other baseline algorithms. The results of the ablation experiments further confirmed the effectiveness of CoLLaRS.

Dynamic heterogeneous federated learning with multi-level prototypes

Federated learning shows promise as a privacy-preserving collaborative learning technique. Existing research mainly focuses on skewing the class distribution across clients. However, most approaches suffer from catastrophic forgetting and classifier shift, mainly when the global distribution of all classes is extremely unbalanced and the data distribution of the client dynamically evolves over time. In this paper, we study the Dynamic Heterogeneous Federated Learning, which addresses the practical scenario where heterogeneous data distributions exist among different clients and dynamic tasks within the client. Accordingly, we propose a novel federated learning framework named Federated Multi-Level Prototypes and design federated multi-level regularizations. To mitigate classifier shift, we construct semantic prototypes to provide fruitful generalization knowledge. To maintain the model stability and consistency convergence, three regularizations are introduced as training losses, i.e., prototype-based regularization, semantic prototype-based regularization, and federated inter-task regularization. Extensive experiments show that the proposed method achieves advanced performance in various settings.