Recent work studies the supervised online continual learning setting where a learner receives a stream of data whose class distribution changes over time. Distinct from other continual learning settings the learner is presented new samples only once and must distinguish between all seen classes. A number of successful methods in this setting focus on storing and replaying a subset of samples alongside incoming data in a computationally efficient manner. One recent proposal ER-AML achieved strong performance in this setting by applying an asymmetric loss based on contrastive learning to the incoming data and replayed data. However, a key ingredient of the proposed method is avoiding contrasts between incoming data and stored data, which makes it impractical for the setting where only one new class is introduced in each phase of the stream. In this work we adapt a recently proposed approach (\textit{BYOL}) from self-supervised learning to the supervised learning setting, unlocking the constraint on contrasts. We then show that supplementing this with additional regularization on class prototypes yields a new method that achieves strong performance in the one-class incremental learning setting and is competitive with the top performing methods in the multi-class incremental setting.
Class-Incremental Learning (CIL) struggles with catastrophic forgetting when learning new knowledge, and Data-Free CIL (DFCIL) is even more challenging without access to the training data of previously learned classes. Though recent DFCIL works introduce techniques such as model inversion to synthesize data for previous classes, they fail to overcome forgetting due to the severe domain gap between the synthetic and real data. To address this issue, this paper proposes relation-guided representation learning (RRL) for DFCIL, dubbed R-DFCIL. In RRL, we introduce relational knowledge distillation to flexibly transfer the structural relation of new data from the old model to the current model. Our RRL-boosted DFCIL can guide the current model to learn representations of new classes better compatible with representations of previous classes, which greatly reduces forgetting while improving plasticity. To avoid the mutual interference between representation and classifier learning, we employ local rather than global classification loss during RRL. After RRL, the classification head is refined with global class-balanced classification loss to address the data imbalance issue as well as learn the decision boundaries between new and previous classes. Extensive experiments on CIFAR100, Tiny-ImageNet200, and ImageNet100 demonstrate that our R-DFCIL significantly surpasses previous approaches and achieves a new state-of-the-art performance for DFCIL. Code is available at https://github.com/jianzhangcs/R-DFCIL
Federated learning (FL) has attracted growing attention via data-private collaborative training on decentralized clients. However, most existing methods unrealistically assume object classes of the overall framework are fixed over time. It makes the global model suffer from significant catastrophic forgetting on old classes in real-world scenarios, where local clients often collect new classes continuously and have very limited storage memory to store old classes. Moreover, new clients with unseen new classes may participate in the FL training, further aggravating the catastrophic forgetting of the global model. To address these challenges, we develop a novel Global-Local Forgetting Compensation (GLFC) model, to learn a global class incremental model for alleviating the catastrophic forgetting from both local and global perspectives. Specifically, to address local forgetting caused by class imbalance at the local clients, we design a class-aware gradient compensation loss and a class-semantic relation distillation loss to balance the forgetting of old classes and distill consistent inter-class relations across tasks. To tackle the global forgetting brought by the non-i.i.d class imbalance across clients, we propose a proxy server that selects the best old global model to assist the local relation distillation. Moreover, a prototype gradient-based communication mechanism is developed to protect privacy. Our model outperforms state-of-the-art methods by 4.4%-15.1% in terms of average accuracy on representative benchmark datasets.
Knowledge distillation (KD) is a substantial strategy for transferring learned knowledge from one neural network model to another. A vast number of methods have been developed for this strategy. While most method designs a more efficient way to facilitate knowledge transfer, less attention has been put on comparing the effect of knowledge sources such as features, logits, and gradients. This work provides a new perspective to motivate a set of knowledge distillation strategies by approximating the classical KL-divergence criteria with different knowledge sources, making a systematic comparison possible in model compression and incremental learning. Our analysis indicates that logits are generally a more efficient knowledge source and suggests that having sufficient feature dimensions is crucial for the model design, providing a practical guideline for effective KD-based transfer learning.
Existing continual learning techniques focus on either task incremental learning (TIL) or class incremental learning (CIL) problem, but not both. CIL and TIL differ mainly in that the task-id is provided for each test sample during testing for TIL, but not provided for CIL. Continual learning methods intended for one problem have limitations on the other problem. This paper proposes a novel unified approach based on out-of-distribution (OOD) detection and task masking, called CLOM, to solve both problems. The key novelty is that each task is trained as an OOD detection model rather than a traditional supervised learning model, and a task mask is trained to protect each task to prevent forgetting. Our evaluation shows that CLOM outperforms existing state-of-the-art baselines by large margins. The average TIL/CIL accuracy of CLOM over six experiments is 87.6/67.9% while that of the best baselines is only 82.4/55.0%.
Learning using privileged information (LUPI) paradigm, which pioneered teacher-student interaction mechanism, makes the learning models use additional information in training stage. This paper is the first to propose an incremental learning algorithm with LUPI paradigm for stochastic configuration network (SCN), named SCN+. This novel algorithm can leverage privileged information into SCN in the training stage, which provides a new method to train SCN. Moreover, the convergences have been studied in this paper. Finally, experimental results indicate that SCN+ indeed performs favorably.
Non-exemplar class-incremental learning is to recognize both the old and new classes when old class samples cannot be saved. It is a challenging task since representation optimization and feature retention can only be achieved under supervision from new classes. To address this problem, we propose a novel self-sustaining representation expansion scheme. Our scheme consists of a structure reorganization strategy that fuses main-branch expansion and side-branch updating to maintain the old features, and a main-branch distillation scheme to transfer the invariant knowledge. Furthermore, a prototype selection mechanism is proposed to enhance the discrimination between the old and new classes by selectively incorporating new samples into the distillation process. Extensive experiments on three benchmarks demonstrate significant incremental performance, outperforming the state-of-the-art methods by a margin of 3%, 3% and 6%, respectively.
Novel classes frequently arise in our dynamically changing world, e.g., new users in the authentication system, and a machine learning model should recognize new classes without forgetting old ones. This scenario becomes more challenging when new class instances are insufficient, which is called few-shot class-incremental learning (FSCIL). Current methods handle incremental learning retrospectively by making the updated model similar to the old one. By contrast, we suggest learning prospectively to prepare for future updates, and propose ForwArd Compatible Training (FACT) for FSCIL. Forward compatibility requires future new classes to be easily incorporated into the current model based on the current stage data, and we seek to realize it by reserving embedding space for future new classes. In detail, we assign virtual prototypes to squeeze the embedding of known classes and reserve for new ones. Besides, we forecast possible new classes and prepare for the updating process. The virtual prototypes allow the model to accept possible updates in the future, which act as proxies scattered among embedding space to build a stronger classifier during inference. FACT efficiently incorporates new classes with forward compatibility and meanwhile resists forgetting of old ones. Extensive experiments validate FACT's state-of-the-art performance. Code is available at: https://github.com/zhoudw-zdw/CVPR22-Fact
Continual learning, also known as lifelong learning, is an emerging research topic that has been attracting increasing interest in the field of machine learning. With human activity recognition (HAR) playing a key role in enabling numerous real-world applications, an essential step towards the long-term deployment of such recognition systems is to extend the activity model to dynamically adapt to changes in people's everyday behavior. Current research in continual learning applied to HAR domain is still under-explored with researchers exploring existing methods developed for computer vision in HAR. Moreover, analysis has so far focused on task-incremental or class-incremental learning paradigms where task boundaries are known. This impedes the applicability of such methods for real-world systems since data is presented in a randomly streaming fashion. To push this field forward, we build on recent advances in the area of continual machine learning and design a lifelong adaptive learning framework using Prototypical Networks, LAPNet-HAR, that processes sensor-based data streams in a task-free data-incremental fashion and mitigates catastrophic forgetting using experience replay and continual prototype adaptation. Online learning is further facilitated using contrastive loss to enforce inter-class separation. LAPNet-HAR is evaluated on 5 publicly available activity datasets in terms of the framework's ability to acquire new information while preserving previous knowledge. Our extensive empirical results demonstrate the effectiveness of LAPNet-HAR in task-free continual learning and uncover useful insights for future challenges.
In this paper, we focus on a new and challenging decentralized machine learning paradigm in which there are continuous inflows of data to be addressed and the data are stored in multiple repositories. We initiate the study of data decentralized class-incremental learning (DCIL) by making the following contributions. Firstly, we formulate the DCIL problem and develop the experimental protocol. Secondly, we introduce a paradigm to create a basic decentralized counterpart of typical (centralized) class-incremental learning approaches, and as a result, establish a benchmark for the DCIL study. Thirdly, we further propose a Decentralized Composite knowledge Incremental Distillation framework (DCID) to transfer knowledge from historical models and multiple local sites to the general model continually. DCID consists of three main components namely local class-incremental learning, collaborated knowledge distillation among local models, and aggregated knowledge distillation from local models to the general one. We comprehensively investigate our DCID framework by using different implementations of the three components. Extensive experimental results demonstrate the effectiveness of our DCID framework. The codes of the baseline methods and the proposed DCIL will be released at https://github.com/zxxxxh/DCIL.
Application of artificial intelligence in medicine brings in highly accurate predictions achieved by complex models, the reasoning of which is hard to interpret. Their generalization ability can be reduced because of the lack of pixel wise annotated images that occurs in frozen section tissue analysis. To partially overcome this gap, this paper explores the approximate explicit feature map (aEFM) transform of low-dimensional data into a low-dimensional subspace in Hilbert space. There, with a modest increase in computational complexity, linear algorithms yield improved performance and keep interpretability. They remain amenable to incremental learning that is not a trivial issue for some nonlinear algorithms. We demonstrate proposed methodology on a very large-scale problem related to intraoperative pixel-wise semantic segmentation and clustering of adenocarcinoma of a colon in a liver. Compared to the results in the input space, logistic classifier achieved statistically significant performance improvements in micro balanced accuracy and F1 score in the amounts of 12.04% and 12.58%, respectively. Support vector machine classifier yielded the increase of 8.04% and 9.41%. For clustering, increases of 0.79% and 0.85% are obtained with ultra large-scale spectral clustering algorithm. Results are supported by a discussion of interpretability using Shapely additive explanation values for predictions of linear classifier in input space and aEFM induced space.
Deep Siamese trackers have recently gained much attention in recent years since they can track visual objects at high speeds. Additionally, adaptive tracking methods, where target samples collected by the tracker are employed for online learning, have achieved state-of-the-art accuracy. However, single object tracking (SOT) remains a challenging task in real-world application due to changes and deformations in a target object's appearance. Learning on all the collected samples may lead to catastrophic forgetting, and thereby corrupt the tracking model. In this paper, SOT is formulated as an online incremental learning problem. A new method is proposed for dynamic sample selection and memory replay, preventing template corruption. In particular, we propose a change detection mechanism to detect gradual changes in object appearance and select the corresponding samples for online adaption. In addition, an entropy-based sample selection strategy is introduced to maintain a diversified auxiliary buffer for memory replay. Our proposed method can be integrated into any object tracking algorithm that leverages online learning for model adaptation. Extensive experiments conducted on the OTB-100, LaSOT, UAV123, and TrackingNet datasets highlight the cost-effectiveness of our method, along with the contribution of its key components. Results indicate that integrating our proposed method into state-of-art adaptive Siamese trackers can increase the potential benefits of a template update strategy, and significantly improve performance.
There exists a large number of datasets for organ segmentation, which are partially annotated and sequentially constructed. A typical dataset is constructed at a certain time by curating medical images and annotating the organs of interest. In other words, new datasets with annotations of new organ categories are built over time. To unleash the potential behind these partially labeled, sequentially-constructed datasets, we propose to incrementally learn a multi-organ segmentation model. In each incremental learning (IL) stage, we lose the access to previous data and annotations, whose knowledge is assumingly captured by the current model, and gain the access to a new dataset with annotations of new organ categories, from which we learn to update the organ segmentation model to include the new organs. While IL is notorious for its `catastrophic forgetting' weakness in the context of natural image analysis, we experimentally discover that such a weakness mostly disappears for CT multi-organ segmentation. To further stabilize the model performance across the IL stages, we introduce a light memory module and some loss functions to restrain the representation of different categories in feature space, aggregating feature representation of the same class and separating feature representation of different classes. Extensive experiments on five open-sourced datasets are conducted to illustrate the effectiveness of our method.
We introduce a memory-based approach to part of speech tagging. Memory-based learning is a form of supervised learning based on similarity-based reasoning. The part of speech tag of a word in a particular context is extrapolated from the most similar cases held in memory. Supervised learning approaches are useful when a tagged corpus is available as an example of the desired output of the tagger. Based on such a corpus, the tagger-generator automatically builds a tagger which is able to tag new text the same way, diminishing development time for the construction of a tagger considerably. Memory-based tagging shares this advantage with other statistical or machine learning approaches. Additional advantages specific to a memory-based approach include (i) the relatively small tagged corpus size sufficient for training, (ii) incremental learning, (iii) explanation capabilities, (iv) flexible integration of information in case representations, (v) its non-parametric nature, (vi) reasonably good results on unknown words without morphological analysis, and (vii) fast learning and tagging. In this paper we show that a large-scale application of the memory-based approach is feasible: we obtain a tagging accuracy that is on a par with that of known statistical approaches, and with attractive space and time complexity properties when using {\em IGTree}, a tree-based formalism for indexing and searching huge case bases.} The use of IGTree has as additional advantage that optimal context size for disambiguation is dynamically computed.
Place recognition is a fundamental component of robotics, and has seen tremendous improvements through the use of deep learning models in recent years. Networks can experience significant drops in performance when deployed in unseen or highly dynamic environments, and require additional training on the collected data. However naively fine-tuning on new training distributions can cause severe degradation of performance on previously visited domains, a phenomenon known as catastrophic forgetting. In this paper we address the problem of incremental learning for point cloud place recognition and introduce InCloud, a structure-aware distillation-based approach which preserves the higher-order structure of the network's embedding space. We introduce several challenging new benchmarks on four popular and large-scale LiDAR datasets (Oxford, MulRan, In-house and KITTI) showing broad improvements in point cloud place recognition performance over a variety of network architectures. To the best of our knowledge, this work is the first to effectively apply incremental learning for point cloud place recognition. Data pre-processing, training and evaluation code for this paper can be found at https://github.com/csiro-robotics/InCloud.
Motivated by recent findings from cognitive neural science, we advocate the use of a dual-level model for concept representations: the embodied level consists of concept-oriented feature representations, and the symbolic level consists of concept graphs. Embodied concept representations are modality specific and exist in the form of feature vectors in a feature space. Symbolic concept representations, on the other hand, are amodal and language specific, and exist in the form of word / knowledge-graph embeddings in a concept / knowledge space. The human conceptual system comprises both embodied representations and symbolic representations, which typically interact to drive conceptual processing. As such, we further advocate the use of dual embodied-symbolic concept representations for deep learning. To demonstrate their usage and value, we discuss two important use cases: embodied-symbolic knowledge distillation for few-shot class incremental learning, and embodied-symbolic fused representation for image-text matching. Dual embodied-symbolic concept representations are the foundation for deep learning and symbolic AI integration. We discuss two important examples of such integration: scene graph generation with knowledge graph bridging, and multimodal knowledge graphs.
The goal of electronic monitoring of longline fishing is to visually monitor the fish catching activities on fishing vessels based on cameras, either for regulatory compliance or catch counting. The previous hierarchical classification method demonstrates efficient fish species identification of catches from longline fishing, where fishes are under severe deformation and self-occlusion during the catching process. Although the hierarchical classification mitigates the laborious efforts of human reviews by providing confidence scores in different hierarchical levels, its performance drops dramatically under the class incremental learning (CIL) scenario. A CIL system should be able to learn about more and more classes over time from a stream of data, i.e., only the training data for a small number of classes have to be present at the beginning and new classes can be added progressively. In this work, we introduce a Hierarchical Class Incremental Learning (HCIL) model, which significantly improves the state-of-the-art hierarchical classification methods under the CIL scenario.
Incremental learning methods can learn new classes continually by distilling knowledge from the last model (as a teacher model) to the current model (as a student model) in the sequentially learning process. However, these methods cannot work for Incremental Implicitly-Refined Classification (IIRC), an incremental learning extension where the incoming classes could have two granularity levels, a superclass label and a subclass label. This is because the previously learned superclass knowledge may be occupied by the subclass knowledge learned sequentially. To solve this problem, we propose a novel Multi-Teacher Knowledge Distillation (MTKD) strategy. To preserve the subclass knowledge, we use the last model as a general teacher to distill the previous knowledge for the student model. To preserve the superclass knowledge, we use the initial model as a superclass teacher to distill the superclass knowledge as the initial model contains abundant superclass knowledge. However, distilling knowledge from two teacher models could result in the student model making some redundant predictions. We further propose a post-processing mechanism, called as Top-k prediction restriction to reduce the redundant predictions. Our experimental results on IIRC-ImageNet120 and IIRC-CIFAR100 show that the proposed method can achieve better classification accuracy compared with existing state-of-the-art methods.
Extensive researches have applied deep neural networks (DNNs) in class incremental learning (Class-IL). As building blocks of DNNs, batch normalization (BN) standardizes intermediate feature maps and has been widely validated to improve training stability and convergence. However, we claim that the direct use of standard BN in Class-IL models is harmful to both the representation learning and the classifier training, thus exacerbating catastrophic forgetting. In this paper we investigate the influence of BN on Class-IL models by illustrating such BN dilemma. We further propose BN Tricks to address the issue by training a better feature extractor while eliminating classification bias. Without inviting extra hyperparameters, we apply BN Tricks to three baseline rehearsal-based methods, ER, DER++ and iCaRL. Through comprehensive experiments conducted on benchmark datasets of Seq-CIFAR-10, Seq-CIFAR-100 and Seq-Tiny-ImageNet, we show that BN Tricks can bring significant performance gains to all adopted baselines, revealing its potential generality along this line of research.
Understanding the interconnected relationships of large-scale information networks like social, scholar and Internet of Things networks is vital for tasks like recommendation and fraud detection. The vast majority of the real-world networks are inherently heterogeneous and dynamic, containing many different types of nodes and edges and can change drastically over time. The dynamicity and heterogeneity make it extremely challenging to reason about the network structure. Unfortunately, existing approaches are inadequate in modeling real-life dynamical networks as they either have strong assumption of a given stochastic process or fail to capture the heterogeneity of network structure, and they all require extensive computational resources. We introduce Lime , a better approach for modeling dynamic and heterogeneous information networks. Lime is designed to extract high-quality network representation with significantly lower memory resources and computational time over the state-of-the-arts. Unlike prior work that uses a vector to encode each network node, we exploit the semantic relationships among network nodes to encode multiple nodes with similar semantics in shared vectors. By using many fewer node vectors, our approach significantly reduces the required memory space for encoding large-scale networks. To effectively trade information sharing for reduced memory footprint, we employ the recursive neural network (RsNN) with carefully designed optimization strategies to explore the node semantics in a novel cuboid space. We then go further by showing, for the first time, how an effective incremental learning approach can be developed – with the help of RsNN, our cuboid structure, and a set of novel optimization techniques – to allow a learning framework to quickly and efficiently adapt to a constantly evolving network. We evaluate Lime by applying it to three representative network-based tasks, node classification, node clustering and anomaly detection, performin...