Generalized Category Discovery (GCD) aims to classify labeled instances from known categories while discovering novel categories from unlabeled data. Despite recent progress in GCD for computer vision, existing GCD approaches largely rely on static final-step representations (in the visual domain), overlooking the temporally evolving nature of time-series data. In this paper, we introduce TGCD, the first framework specifically designed for GCD in time-series data. TGCD leverages both the dynamics of latent representations and the heterogeneity of predictions across multiple temporal segments to disover unknown (i.e., novel) categories, based on a pre-trained time-series foundation model. We propose a unified learning objective for TGCD that integrates the following three components: (i) a Stochastic Temporal Segment Dropout (STeSD) objective that regularizes the model by selectively penalizing high-entropy segments to encourage confident predictions on uncertain regions of the time-series, and (ii) a Known–Unknown Temporal Discriminability (KUTD) objective that promotes representational separation between known and unknown categories within unlabeled data and (iii) a margin-aware classification objective to improve generalization. Empirical evaluation on six multivariate time-series data sets demonstrates that the TGCD substantially outperforms existing GCD methods, particularly in discovering unknown categories. We further conduct ablation studies to highlight the individual contributions of each component. Additionally, we provide the first comprehensive benchmarking of recent GCD approaches on time-series data, revealing the limitations of naive transfer and underscoring the benefits of temporal modeling.

Deep learning models are increasingly used for making predictions based on clinical time-series data, but model generalization remains a challenge. Continual learning approaches, which preserve representations while learning new distributions, are suitable for addressing this challenge. We propose Continual Bayesian Long Short Term Memory (C-BLSTM), a continual learning algorithm based on the Bayesian LSTM model for domain incremental learning. C-BLSTM continually learns a sequence of tasks by combining architectural pruning, variational inference-based regularization, and coreset replay strategies. In extensive experiments on two public electronic medical record datasets for mortality prediction, we show that C-BLSTM outperforms many state-of-the-art continual learning approaches. Further, we apply the C-BLSTM to two real-world clinical time series datasets for prediction of readmission risk in patients with heart failure and glycated haemoglobin outcomes in patients with type 2 diabetes. First, we show that these datasets exhibit domain incremental characteristics with significant drifts in their marginal distributions and moderate drifts in their conditional distributions. Then, we demonstrate that the C-BLSTM improves generalization in five diverse real-world scenarios spanning temporal, site, device, case mix, and ethnicity shifts, both in terms of performance and reliability of predictions.

Romain Meunier, Farah Benamara, Véronique Moriceau, Zhongzheng Qiao, Savitha Ramasamy. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2025.

The data privacy constraint in online continual learning (OCL), where the data can be seen only once, complicates the catastrophic forgetting problem in streaming data. A common approach applied by the current SOTAs in OCL is with the use of memory saving exemplars or features from previous classes to be replayed in the current task. On the other hand, the prompt-based approach performs excellently in continual learning but with the cost of a growing number of trainable parameters. The first approach may not be applicable in practice due to data openness policy, while the second approach has the issue of throughput associated with the streaming data. In this study, we propose a novel prompt-based method for online continual learning that includes 4 main components: (1) single light-weight prompt generator as a general knowledge, (2) trainable scaler-and-shifter as specific knowledge, (3) pre-trained model (PTM) generalization preserving, and (4) hard-soft updates mechanism. Our proposed method achieves significantly higher performance than the current SOTAs in CIFAR100, ImageNet-R, ImageNet-A, and CUB dataset. Our complexity analysis shows that our method requires a relatively smaller number of parameters and achieves moderate training time, inference time, and throughput. For further study, the source code of our method is available at https://github.com/anwarmaxsum/PROL.

The prompt-based approach has demonstrated its success for continual learning problems. However, it still suffers from catastrophic forgetting due to inter-task vector similarity and unfitted new components of previously learned tasks. On the other hand, the language-guided approach falls short of its full potential due to minimum utilized knowledge and participation in the prompt tuning process. To correct this problem, we propose a novel prompt-based structure and algorithm that incorporate 4 key concepts (1) language as input for prompt generation (2) task-wise generators (3) limiting matching descriptors search space via soft task-id prediction (4) generated prompt as auxiliary data. Our experimental analysis shows the superiority of our method to existing SOTAs in CIFAR100, ImageNet-R, and CUB datasets with significant margins i.e. up to 30% final average accuracy, 24% cumulative average accuracy, 8% final forgetting measure, and 7% cumulative forgetting measure. Our historical analysis confirms our method successfully maintains the stability-plasticity trade-off in every task. Our robustness analysis shows the proposed method consistently achieves high performances in various prompt lengths, layer depths, and number of generators per task compared to the SOTAs. We provide a comprehensive theoretical analysis, and complete numerical results in appendix sections. The method code is available in https://github.com/anwarmaxsum/LEAPGEN for further study.

Time series foundation models (TSFMs) demonstrate impressive zero-shot performance for time series forecasting. However, an important yet underexplored challenge is how to effectively finetune TSFMs on specific downstream tasks. While naive finetuning can yield performance gains, we argue that it falls short of fully leveraging TSFMs' capabilities, often resulting in overfitting and suboptimal performance. Given the diverse temporal patterns across sampling scales and the inherent multi-scale forecasting capabilities of TSFMs, we adopt a causal perspective to analyze finetuning process, through which we highlight the critical importance of explicitly modeling multiple scales and reveal the shortcomings of naive approaches. Focusing on encoder-based TSFMs, we propose Multiscale finetuning (MSFT), a simple yet general framework that explicitly integrates multi-scale modeling into the finetuning process. Experimental results on three different backbones (Moirai, Moment and Units) demonstrate that TSFMs finetuned with MSFT not only outperform naive and typical parameter efficient finetuning methods but also surpass state-of-the-art deep learning methods. Codes are available at https://github.com/zqiao11/MSFT.

As vision-language models advance, addressing the Zero-Shot Learning (ZSL) problem in the open world becomes increasingly crucial. Specifically, a robust model must handle three types of samples during inference: seen classes with visual and semantic information provided in training, unseen classes with only the semantic information in training, and unknown samples with no prior information from training. Existing methods either handle seen and unseen classes together (ZSL) or seen and unknown classes (known as Open-Set Recognition, OSR). However, none addresses the simultaneous handling of all three, which we term Open-Set Zero-Shot Learning (OZSL). To address this problem, we propose a two-stage approach for OZSL that recognizes seen, unseen, and unknown samples. The first stage classifies samples as either seen or not, while the second stage distinguishes unseen from unknown. Furthermore, we introduce a cross-stage knowledge transfer mechanism that leverages semantic relationships between seen and unseen classes to enhance learning in the second stage. Extensive experiments demonstrate the efficacy of the proposed approach compared to naívely combining existing ZSL and OSR methods. The code is available at https://github.com/smufang/OZSL.

Spatio-temporal graph neural networks have proven efficacy in capturing complex dependencies for urban computing tasks such as forecasting and kriging. Yet, their performance is constrained by the reliance on extensive data for training on a specific task, thereby limiting their adaptability to new urban domains with varied task demands. Although transfer learning has been proposed to remedy this problem by leveraging knowledge across domains, the cross-task generalization still remains under-explored in spatio-temporal graph transfer learning due to the lack of a unified framework. To bridge the gap, we propose Spatio-Temporal Graph Prompting (STGP), a prompt-based framework capable of adapting to multi-diverse tasks in a data-scarce domain. Specifically, we first unify different tasks into a single template and introduce a task-agnostic network architecture that aligns with this template. This approach enables capturing dependencies shared across tasks. Furthermore, we employ learnable prompts to achieve domain and task transfer in a two-stage prompting pipeline, facilitating the prompts to effectively capture domain knowledge and task-specific properties. Our extensive experiments demonstrate that STGP outperforms state-of-the-art baselines in three tasks-forecasting, kriging, and extrapolation-achieving an improvement of up to 10.7%.

Federated Class Incremental Learning (FCIL) is a new direction in continual learning (CL) for addressing catastrophic forgetting and non-IID data distribution simultaneously. Existing FCIL methods call for high communication costs and exemplars from previous classes. We propose a novel rehearsal-free method for FCIL named prototypes-injected prompt (PIP) that involves 3 main ideas: a) prototype injection on prompt learning, b) prototype augmentation, and c) weighted Gaussian aggregation on the server side. Our experiment result shows that the proposed method outperforms the current state of the arts (SOTAs) with a significant improvement (up to 33%) in CIFAR100, MiniImageNet, and TinyImageNet datasets. Our extensive analysis demonstrates the robustness of PIP in different task sizes, and the advantage of requiring smaller participating local clients, and smaller global rounds. For further study, source codes of PIP, baseline, and experimental logs are shared publicly in https://github.com/anwarmaxsum/PIP.

In autonomous and mobile robotics, a principal challenge is resilient real-time environmental perception, particularly in situations characterized by unknown and dynamic elements, as exemplified in the context of autonomous drone racing. This study introduces a perception technique for detecting drone racing gates under illumination variations, which is common during high-speed drone flights. The proposed technique relies upon a lightweight neural network backbone augmented with capabilities for continual learning. The envisaged approach amalgamates predictions of the gates' positional coordinates, distance, and orientation, encapsulating them into a cohesive pose tuple. A comprehensive number of tests serve to underscore the efficacy of this approach in confronting diverse and challenging scenarios, specifically those involving variable lighting conditions. The proposed methodology exhibits notable robustness in the face of illumination variations, thereby substantiating its effectiveness.

Real-world environments are inherently non-stationary, frequently introducing new classes over time. This is especially common in time series classification, such as the emergence of new disease classification in healthcare or the addition of new activities in human activity recognition. In such cases, a learning system is required to assimilate novel classes effectively while avoiding catastrophic forgetting of the old ones, which gives rise to the Class-incremental Learning (CIL) problem. However, despite the encouraging progress in the image and language domains, CIL for time series data remains relatively understudied. Existing studies suffer from inconsistent experimental designs, necessitating a comprehensive evaluation and benchmarking of methods across a wide range of datasets. To this end, we first present an overview of the Time Series Class-incremental Learning (TSCIL) problem, highlight its unique challenges, and cover the advanced methodologies. Further, based on standardized settings, we develop a unified experimental framework that supports the rapid development of new algorithms, easy integration of new datasets, and standardization of the evaluation process. Using this framework, we conduct a comprehensive evaluation of various generic and time-series-specific CIL methods in both standard and privacy-sensitive scenarios. Our extensive experiments not only provide a standard baseline to support future research but also shed light on the impact of various design factors such as normalization layers or memory budget thresholds. Codes are available at https://github.com/zqiao11/TSCIL.

EdTech platforms continuously refresh their database with new questions and concepts with evolving course syllabus. The state-of-the-art knowledge tracing models are unable to adapt to these changes, as the size of the question embedding layers is typically fixed. In this work, we propose an incremental learning algorithm for knowledge tracing that is capable of adapting itself to growing pool of concepts and questions, through its architectural adaptation and regularization strategies. The algorithm, referred as, "Architectural adaptation and Regularization of Attention network for Incremental Knowledge Tracing (ARAIKT)", is capable of adapting the embeddings with increasing concepts and question bank, while preserving representations of the previous concepts and question banks. Furthermore, they are robust to distributional drifts in the data, and are capable of preserving privacy of data across study centers and EdTech platforms. We demonstrate the effectiveness of the ARAIKT by evaluating its performance on subsets of study centers/academic years within ASSISTment2009 and ASSISTment2017 data sets, respectively.

This article presents a new approach for providing an interpretation for a spiking neural network classifier by transforming it to a multiclass additive model. The spiking classifier is a multiclass synaptic efficacy function-based leaky-integrate-fire neuron (Mc-SEFRON) classifier. As a first step, the SEFRON classifier for binary classification is extended to handle multiclass classification problems. Next, a new method is presented to transform the temporally distributed weights in a fully trained Mc-SEFRON classifier to shape functions in the feature space. A composite of these shape functions results in an interpretable classifier, namely, a directly interpretable multiclass additive model (DIMA). The interpretations of DIMA are also demonstrated using the multiclass Iris dataset. Further, the performances of both the Mc-SEFRON and DIMA classifiers are evaluated on ten benchmark datasets from the UCI machine learning repository and compared with the other state-of-the-art spiking neural classifiers. The performance study results show that Mc-SEFRON produces similar or better performances than other spiking neural classifiers with an added benefit of interpretability through DIMA. Furthermore, the minor differences in accuracies between Mc-SEFRON and DIMA indicate the reliability of the DIMA classifier. Finally, the Mc-SEFRON and DIMA are tested on three real-world credit scoring problems, and their performances are compared with state-of-the-art results using machine learning methods. The results clearly indicate that DIMA improves the classification accuracy by up to 12% over other interpretable classifiers indicating a better quality of interpretations on the highly imbalanced credit scoring datasets.

Truong Do, Le Khiem, Quang Pham, TrungTin Nguyen, Thanh-Nam Doan, Binh Nguyen, Chenghao Liu, Savitha Ramasamy, Xiaoli Li, Steven Hoi. Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing. 2023.

This paper tackles the out-of-distribution (OOD) detection problem for natural language classifiers. While the previous OOD detection methods require large-scale in-distribution (ID) training data, we attack this problem from the few-shot perspective in an unsupervised manner where the training relies on only a few samples from ID data. First, we develop various baselines for Few-shot OOD (FSOOD) detection in text classification based on the three well-known few-shot learning approaches (well-explored in the vision domain), i.e., meta-learning, metric learning, and data augmentation (DA). Then, we introduce the concept of demonstration-based data augmentation with meta and metric-learning approaches to reap the combined benefit of both approaches. A pre-trained transformer is fine-tuned on a few available ID samples in all developed methods. In tandem with this fine-tuning, an OOD detector is fitted over the ID training samples to reject the data from the unknown classes using two kinds of distance metrics, namely Mahalanobis distance and Cosine similarity. At last, we present an extensive evaluation of three ID datasets and three OOD datasets. We also perform an ablation study to analyze the impact of various components of our method.

Class-incremental learning (CIL) on multivariate time series (MTS) is an important yet understudied problem. Based on practical privacy-sensitive circumstances, we propose a novel distillation-based strategy using a single-headed classifier without saving historical samples. We propose to exploit Soft-Dynamic Time Warping (Soft-DTW) for knowledge distillation, which aligns the feature maps along the temporal dimension before calculating the discrepancy. Compared with Euclidean distance, Soft-DTW shows its advantages in overcoming catastrophic forgetting and balancing the stability-plasticity dilemma. We construct two novel MTS-CIL benchmarks for comprehensive experiments. Combined with a prototype augmentation strategy, our framework demonstrates significant superiority over other prominent exemplar-free algorithms.

This work presents a novel approach which integrates deep learning, online learning and continual learning paradigms for adaptive control for robotic systems. Deep learning allows generalising knowledge about the robot, while online learning can adapt to variable operating conditions, and continual learning enables remembering previous knowledge. The proposed method approximates the inverse dynamics of the robot, which is formulated as a regression problem. With a minimum knowledge of the robot's dynamics, the proposed method shows its capability to reduce tracking errors online by continuously learning and compensating for internal and external changing conditions. Furthermore, the simulation results show that the proposed approach with online continual learning improves the control performance of ground and aerial mobile robots.

Learning from irregularly sampled, streaming, multi-variate time-series data with many missing values is a very challenging task. In this paper, we propose a Bayesian Continual Imputation and Prediction for Time-series Data (B-CIPIT), for learning from a sequence of time-series tasks. First, we develop a Bayesian LSTM based continual learning algorithm, which is capable of learning continually from a sequence of multi-variate time-series tasks, without catastrophically forgetting any representations. Second, we impute missing values in these time-series sequences, in a continual learning setting. We demonstrate and evaluate the robustness of the proposed algorithm on two real-world clinical time-series data sets, namely MIMIC-III [1] and PhysioNet Challenge 2012 [2]. Performance study results show the superiority of the proposed learning algorithm.

We propose an online continual learning algorithm based on an enhanced Random Vector Functional Link Network (OCL-eRVFL), that learns a sequence of tasks continually, where each task is defined by streaming data with each sample arriving once and only once. As data for a new task in domain incremental or class incremental setting streams in, the output weights of an eRVFL is updated through Recursive least squares, such that the representations for the past tasks are not catastrophically forgotten. As the recursive least square update is based only on the currently streaming sample, samples are not stored. Hence, unlike state-of-the-art OCL that avoid catastrophic forgetting through memory replay of samples from past task, the proposed OCL-eRVFL needs no extra memory. The proposed OCL-eRVFL is evaluated on streaming split CIFAR10, split CIFAR100 and split CIFAR10/100 image classification data sets within a class and domain incremental setting. Performance results show that the proposed OCL-eRVFL efficiently learns a sequence of tasks with streaming data, without additional memory expense.

Knowledge Tracing is the prediction of the future performance of a learner, given the past performance. The existing knowledge tracing models represent the training data and does not generalize when there is a drift in the data distribution. We first empirically demonstrate an evolving Knowledge Tracing (eKT) scenario with distinct distribution of learner performances and diversity of questions from similar concepts. Next, we empirically characterize drift in the data and propose a task agnostic incremental context aware attentive knowledge tracing (iAKT) approach to learn incrementally from the eKT. The iAKT regularizes representations to learn from diverse learner performance distributions. Finally, we evaluate the ability of the proposed iAKT for knowledge tracing and study the effect of various regularization strategies on ranking difficulty of questions, using the ASSISTments 2017 data set. Performance results show that the iAKT adapts its representations to drift in data characteristics, while iAKT with EWC regularizer is better at ranking the difficulty of questions.

Robustness of learning algorithms remains an important problem to be solved from both the perspective of adversarial attacks and improving generalization. In this work, we investigate the robustness of biologically inspired Hebbian learning algorithm in depth. We find that Hebbian learning based algorithms outperform conventional learning algorithms like CNNs by a huge margin of upto 18% on the CIFAR-10 dataset under the addition of noise. We highlight that an important reason for this is the underlying representations that are being learnt by the learning algorithms. Specifically, we find that the Hebbian method learns the most robust representations compared to other methods that helps it to generalize better. We also conduct ablations on the Hebbian network and showcase that robustness of the model drops by upto 16% on the CIFAR-10 dataset if the representation capacity of the network is deteriorated. Hence, we find that the representations learnt play an important role in the resultant robustness of the models. We conduct experiments on multiple datasets and show that the results hold on all the datasets and at various noise levels.

Continual learning aims at learning a sequence of tasks without forgetting any task. While most of the existing literature in continual learning is aimed at class incremental learning in a supervised setting, there is an enormous potential for unsupervised continual learning using generative models. This paper proposes a combination of architectural pruning and neuron addition in generative variational models toward unsupervised generative continual learning (UGCL). Evaluations on standard benchmark data sets demonstrate the superior generative ability of the proposed method.

Deep neural networks model data for a task or a sequence of tasks, where the knowledge extracted from the data is encoded in the parameters and representations of the network. Extraction and utilization of these representations is vital when data is no longer available in the future, especially in a continual learning scenario. We introduce flashcards, which are visual representations that capture the encoded knowledge of a network as a recursive function of some predefined random image patterns. In a continual learning scenario, flashcards help to prevent catastrophic forgetting by consolidating the knowledge of all the previous tasks. Flashcards are required to be constructed only before learning the subsequent task, hence, they are independent of the number of tasks trained before, making them task agnostic. We demonstrate the efficacy of flashcards in capturing learned knowledge representation (as an alternative to the original data), and empirically validate on a variety of continual learning tasks: reconstruction, denoising, and task-incremental classification, using several heterogeneous (varying background and complexity) benchmark datasets. Experimental evidence indicates that: (i) flashcards as a replay strategy is task agnostic, (ii) performs better than generative replay, and (iii) is on par with episodic replay without additional memory overhead.

Backpropagation has revolutionized neural network training however, its biological plausibility remains questionable. Hebbian learning, a completely unsupervised and feedback free learning technique is a strong contender for a biologically plausible alternative. However, so far, it has neither achieved high accuracy performance vs. backprop, nor is the training procedure simple. In this work, we introduce a new Hebbian learning based neural network, called HebbNet. At the heart of HebbNet is an improved Hebbian approach that includes an updated activation threshold and gradient sparsity to the first principles of Hebbian learning. These enable an efficiently performing Hebbian approach with a simple training procedure. Further to this, the improved Hebbian rule also improves training dynamics by reducing the number of training epochs from 1500 to 200 and making training a one-step process from a two-step process. We also reduce heuristics by reducing hyper-parameters from 5 to 1, and number of search runs for hyper-parameter tuning from 12,600 to 13. Notwithstanding this, HebbNet still achieves strong test performance on MNIST and CIFAR-10 datasets vs. state-of-the-art.

Continual learning (CL) aims to learn new tasks by forward transfer of information learnt from previous tasks and without forgetting them. In task incremental CL, task information is vital during both strategy development and inference. Providing such partial knowledge about the test sample demands additional complexity and may become intractable, especially when the sample source is ambiguous. In this work, we design a task-agnostic approach that uses base-child hybrid setup to incrementally learn tasks while mitigating forgetting. Multiple base classifiers guided by reference points learn new tasks and this information is distilled via feature space induced sampling strategy. A central child classifier consolidates information across tasks and infers the task identifier automatically. Experimental results on standard datasets show that the proposed approach outperforms the various state-of-the-art regularization and replay CL algorithms in terms of accuracy, by 50% and 7% with homogeneous and heterogeneous tasks, respectively, in task-agnostic scenarios.

This paper presents a complex-valued interval type-2 neuro-fuzzy inference system (CIT2FIS) and derive its metacognitive projection-based learning (PBL) algorithm. Metacognitive CIT2FIS (Mc-CIT2FIS) consists of a CIT2FIS, which realizes Takagi-Sugeno-Kang type inference mechanism, as its cognitive component. A PBL with self-regulation is its metacognitive component. The rules of CIT2FIS employ interval type-\(2~q\) -Gaussian membership functions that can represent different radial basis functions for different values of \(q\) . As each sample is presented to the network, the metacognitive component monitors the hinge-loss error and class-specific knowledge potential of the current sample to efficiently decide on what-to-learn, when-to-learn, and how-to-learn it. When a new rule is added or existing rules are updated, the optimal parameters of CIT2FIS corresponding to the minimum of the hinge-loss error function are computed using a PBL algorithm derived using the Wirtinger calculus. The performance of Mc-CIT2FIS is evaluated on a set of benchmark real-valued classification problems from the UCI machine learning repository. A circular transformation is used to convert the real-valued features to the complex-valued features in these problems. The performance comparison and statistical study clearly show the superior classification ability of Mc-CIT2FIS. Finally, the proposed complex-valued network is used to solve a practical human action recognition problem that is represented by complex-valued optical flow-based feature set, and a human emotion recognition problem represented using complex-valued Gabor filter-based features. The performance results on these problems substantiate the superior classification ability of Mc-CIT2FIS.

This paper presents a fast learning fully complex-valued classifier to solve real-valued classification problems, called the `Fast Learning Complex-valued Neural Classifier' (FLCNC). The FLCNC is a single hidden layer network with a non-linear, real to complex transformed input layer, a hidden layer with a fully complex activation function and a linear output layer. The neurons in the input layer convert the real-valued input features to the Complex domain using an unique non-linear transformation. At the hidden layer, the complex-valued transformed input features are mapped onto a higher dimensional Complex plane using a fully complex-valued activation function of the type of `sech'. The parameters of the input and hidden neurons of the FLCNC are chosen randomly and the output parameters are estimated analytically which makes the FLCNC to perform fast classification. Moreover, the unique nonlinear input transformation and the orthogonal decision boundaries of the complex-valued neural network help the FLCNC to perform accurate classification. Performance of the FLCNC is demonstrated using a set of multi-category and binary real valued classification problems with both balanced and unbalanced data sets from the UCI machine learning repository. Performance comparison with existing complex-valued and real-valued classifiers show the superior classification performance of the FLCNC.