Overfitting in deep neural networks occurs less frequently than expected. This is a puzzling observation, as theory predicts that greater model capacity should eventually lead to overfitting -- yet this is rarely seen in practice. But what if overfitting does occur, not globally, but in specific sub-regions of the data space? In this work, we introduce a novel score that measures the forgetting rate of deep models on validation data, capturing what we term local overfitting: a performance degradation confined to certain regions of the input space. We demonstrate that local overfitting can arise even without conventional overfitting, and is closely linked to the double descent phenomenon. Building on these insights, we introduce a two-stage approach that leverages the training history of a single model to recover and retain forgotten knowledge: first, by aggregating checkpoints into an ensemble, and then by distilling it into a single model of the original size, thus enhancing performance without added inference cost. Extensive experiments across multiple datasets, modern architectures, and training regimes validate the effectiveness of our approach. Notably, in the presence of label noise, our method -- Knowledge Fusion followed by Knowledge Distillation -- outperforms both the original model and independently trained ensembles, achieving a rare win-win scenario: reduced training and inference complexity.

Deep learning models may converge to suboptimal solutions despite strong validation accuracy, masking an optimization failure we term Trajectory Deviation. This is because as training proceeds, models can abandon high generalization states for specific data sub-populations, thus discarding previously learned latent features without triggering classical overfitting signals. To address this problem we introduce VISTA, an online self-distillation framework that enforces consistency along the optimization trajectory. Using a validation-informed Marginal Coverage score, VISTA identifies expert anchors, which are earlier model states that retain specialized competence over distinct data regions. A coverage-weighted ensemble of these anchors is integrated online during training, regularizing the loss landscape and preserving mastered knowledge. When evaluated across multiple benchmarks, VISTA demonstrates improved robustness and generalization over standard training and prior self-distillation methods, while a lightweight implementation reduces storage overhead by 90

Catastrophic forgetting remains a key challenge in Continual Learning (CL). In replay-based CL with severe memory constraints, performance critically depends on the sample selection strategy for the replay buffer. Most existing approaches construct memory buffers using embeddings learned under supervised objectives. However, class-agnostic, self-supervised representations often encode rich, class-relevant semantics that are overlooked. We propose a new method, Multiple Embedding Replay Selection, MERS, which replaces the buffer selection module with a graph-based approach that integrates both supervised and self-supervised embeddings. Empirical results show consistent improvements over SOTA selection strategies across a range of continual learning algorithms, with particularly strong gains in low-memory regimes. On CIFAR-100 and TinyImageNet, MERS outperforms single-embedding baselines without adding model parameters or increasing replay volume, making it a practical, drop-in enhancement for replay-based continual learning.

Federated Active Learning (FAL) addresses the dual challenges of data privacy and label scarcity, where the absence of a global data view introduces additional hurdles for coordinated query selection. We study cross-silo FAL in the low-budget regime, where annotation decisions are most critical. We characterize, both theoretically and empirically, a heterogeneity reversal: in low-budget settings, homogeneous (IID) data requires stronger coordination to avoid redundant queries, whereas heterogeneous data naturally promotes diversity; this trend reverses at higher budgets. Thus, in contrast to the standard federated learning (FL) narrative where heterogeneity is a primary challenge, we show that IID settings are more challenging for query selection in FAL. Motivated by these findings, we propose a new FAL framework that utilizes federated representation learning to align client data in a shared embedding space. This enables the server to perform globally coordinated active selection over optionally obfuscated client embeddings, while annotation remains local to each client. Although our framework operates in the more challenging low-budget regime, it achieves performance that surpasses existing FAL methods even when they are given substantially larger annotation budgets, demonstrating the value of centralized coordination under privacy constraints.

Active Learning (AL) aims to reduce annotation costs by strategically selecting the most informative samples for labeling. However, most active learning methods struggle in the low-budget regime where only a few labeled examples are available. This issue becomes even more pronounced when annotators provide noisy labels. A common AL approach for the low- and mid-budget regimes focuses on maximizing the coverage of the labeled set across the entire dataset. We propose a novel framework called Noise-Aware Active Sampling (NAS) that extends existing greedy, coverage-based active learning strategies to handle noisy annotations. NAS identifies regions that remain uncovered due to the selection of noisy representatives and enables resampling from these areas. We introduce a simple yet effective noise filtering approach suitable for the low-budget regime, which leverages the inner mechanism of NAS and can be applied for noise filtering before model training. On multiple computer vision benchmarks, including CIFAR100 and ImageNet subsets, NAS significantly improves performance for standard active learning methods across different noise types and rates.

The infrequent occurrence of overfitting in deep neural networks is perplexing: contrary to theoretical expectations, increasing model size often enhances performance in practice. But what if overfitting does occur, though restricted to specific sub-regions of the data space? In this work, we propose a novel score that captures the forgetting rate of deep models on validation data. We posit that this score quantifies local overfitting: a decline in performance confined to certain regions of the data space. We then show empirically that local overfitting occurs regardless of the presence of traditional overfitting. Using the framework of deep over-parametrized linear models, we offer a certain theoretical characterization of forgotten knowledge, and show that it correlates with knowledge forgotten by real deep models. Finally, we devise a new ensemble method that aims to recover forgotten knowledge, relying solely on the training history of a single network. When combined with knowledge distillation, this method will enhance the performance of a trained model without adding inference costs. Extensive empirical evaluations demonstrate the efficacy of our method across multiple datasets, contemporary neural network architectures, and training protocols.

Deep Active Learning (AL) techniques can be effective in reducing annotation costs for training deep models. However, their effectiveness in low- and high-budget scenarios seems to require different strategies, and achieving optimal results across varying budget scenarios remains a challenge. In this study, we introduce Dynamic Coverage & Margin mix (DCoM), a novel active learning approach designed to bridge this gap. Unlike existing strategies, DCoM dynamically adjusts its strategy, considering the competence of the current model. Through theoretical analysis and empirical evaluations on diverse datasets, including challenging computer vision tasks, we demonstrate DCoM's ability to overcome the cold start problem and consistently improve results across different budgetary constraints. Thus DCoM achieves state-of-the-art performance in both low- and high-budget regimes.

Deep neural networks have become the method of choice for solving many classification tasks, largely because they can fit very complex functions defined over raw data. The downside of such powerful learners is the danger of overfit. In this paper, we introduce a novel ensemble classifier for deep networks that effectively overcomes overfitting by combining models generated at specific intermediate epochs during training. Our method allows for the incorporation of useful knowledge obtained by the models during the overfitting phase without deterioration of the general performance, which is usually missed when early stopping is used. To motivate this approach, we begin with the theoretical analysis of a regression model, whose prediction - that the variance among classifiers increases when overfit occurs - is demonstrated empirically in deep networks in common use. Guided by these results, we construct a new ensemble-based prediction method, where the prediction is determined by the class that attains the most consensual prediction throughout the training epochs. Using multiple image and text classification datasets, we show that when regular ensembles suffer from overfit, our method eliminates the harmful reduction in generalization due to overfit, and often even surpasses the performance obtained by early stopping. Our method is easy to implement and can be integrated with any training scheme and architecture, without additional prior knowledge beyond the training set. It is thus a practical and useful tool to overcome overfit.

Continual Learning is an unresolved challenge, whose relevance increases when considering modern applications. Unlike the human brain, trained deep neural networks suffer from a phenomenon called catastrophic forgetting, wherein they progressively lose previously acquired knowledge upon learning new tasks. To mitigate this problem, numerous methods have been developed, many relying on the replay of past exemplars during new task training. However, as the memory allocated for replay decreases, the effectiveness of these approaches diminishes. On the other hand, maintaining a large memory for the purpose of replay is inefficient and often impractical. Here we introduce TEAL, a novel approach to populate the memory with exemplars, that can be integrated with various experience-replay methods and significantly enhance their performance with small memory buffers. We show that TEAL enhances the average accuracy of existing class-incremental methods and outperforms other selection strategies, achieving state-of-the-art performance even with small memory buffers of 1-3 exemplars per class in the final task. This confirms our initial hypothesis that when memory is scarce, it is best to prioritize the most typical data. Code is available at this https URL: https://github.com/shahariel/TEAL.

In the domain of semi-supervised learning (SSL), the conventional approach involves training a learner with a limited amount of labeled data alongside a substantial volume of unlabeled data, both drawn from the same underlying distribution. However, for deep learning models, this standard practice may not yield optimal results. In this research, we propose an alternative perspective, suggesting that distributions that are more readily separable could offer superior benefits to the learner as compared to the original distribution. To achieve this, we present PruneSSL, a practical technique for selectively removing examples from the original unlabeled dataset to enhance its separability. We present an empirical study, showing that although PruneSSL reduces the quantity of available training data for the learner, it significantly improves the performance of various competitive SSL algorithms, thereby achieving state-of-the-art results across several image classification tasks.

In supervised Active Learning (AL), the learner can manipulate the labeled training set by choosing examples to be labeled by an oracle. The size of the labeled set is termed budget. Recent years have seen significant progress in this domain in the context of deep active learning. In particular, it has been shown that in general, different families of AL strategies are suitable for high and low budgets. Here we address for the first time the problem of deciding which family of strategies is most suitable for a given budget in a given task. We start from the theoretical analysis of a mixture model, which motivates a computational approach to decide on the most suitable family of methods for the task and budget at hand. We then propose a practical decision algorithm, which determines what family of strategies should be preferred. Using this algorithm, we introduce MiSAL - a mixed strategy active learning algorithm. MiSAL combines AL strategies from different families, resulting in a method that fits all budgets. We support the analysis by an empirical study, showing the superiority of our method when dealing with image datasets.

In the domain of Active Learning (AL), a learner actively selects which unlabeled examples to seek labels from an oracle, while operating within predefined budget constraints. Importantly, it has been recently shown that distinct query strategies are better suited for different conditions and budgetary constraints. In practice, the determination of the most appropriate AL strategy for a given situation remains an open problem. To tackle this challenge, we propose a practical derivative-based method that dynamically identifies the best strategy for a given budget. Intuitive motivation for our approach is provided by the theoretical analysis of a simplified scenario. We then introduce a method to dynamically select an AL strategy, which takes into account the unique characteristics of the problem and the available budget. Empirical results showcase the effectiveness of our approach across diverse budgets and computer vision tasks.

The infrequent occurrence of overfit in deep neural networks is perplexing. On the one hand, theory predicts that as models get larger they should eventually become too specialized for a specific training set, with ensuing decrease in generalization. In contrast, empirical results in image classification indicate that increasing the training time of deep models or using bigger models almost never hurts generalization. Is it because the way we measure overfit is too limited? Here, we introduce a novel score for quantifying overfit, which monitors the forgetting rate of deep models on validation data. Presumably, this score indicates that even while generalization improves overall, there are certain regions of the data space where it deteriorates. When thus measured, we show that overfit can occur with and without a decrease in validation accuracy, and may be more common than previously appreciated. This observation may help to clarify the aforementioned confusing picture. We use our observations to construct a new ensemble method, based solely on the training history of a single network, which provides significant improvement in performance without any additional cost in training time. An extensive empirical evaluation with modern deep models shows our method's utility on multiple datasets, neural networks architectures and training schemes, both when training from scratch and when using pre-trained networks in transfer learning. Notably, our method outperforms comparable methods while being easier to implement and use, and further improves the performance of competitive networks on Imagenet by 1%.

Semi-Supervised Learning (SSL) is a framework that utilizes both labeled and unlabeled data to enhance model performance. Conventional SSL methods operate under the assumption that labeled and unlabeled data share the same label space. However, in practical real-world scenarios, especially when the labeled training dataset is limited in size, some classes may be totally absent from the labeled set. To address this broader context, we propose a general approach to augment existing SSL methods, enabling them to effectively handle situations where certain classes are missing. This is achieved by introducing an additional term into their objective function, which penalizes the KL-divergence between the probability vectors of the true class frequencies and the inferred class frequencies. Our experimental results reveal significant improvements in accuracy when compared to state-of-the-art SSL, open-set SSL, and open-world SSL methods. We conducted these experiments on two benchmark image classification datasets, CIFAR-100 and STL-10, with the most remarkable improvements observed when the labeled data is severely limited, with only a few labeled examples per class

Deep neural networks have incredible capacity and expressibility, and can seemingly memorize any training set. This introduces a problem when training in the presence of noisy labels, as the noisy examples cannot be distinguished from clean examples by the end of training. Recent research has dealt with this challenge by utilizing the fact that deep networks seem to memorize clean examples much earlier than noisy examples. Here we report a new empirical result: for each example, when looking at the time it has been memorized by each model in an ensemble of networks, the diversity seen in noisy examples is much larger than the clean examples. We use this observation to develop a new method for noisy labels filtration. The method is based on a statistics of the data, which captures the differences in ensemble learning dynamics between clean and noisy data. We test our method on three tasks: (i) noise amount estimation; (ii) noise filtration; (iii) supervised classification. We show that our method improves over existing baselines in all three tasks using a variety of datasets, noise models, and noise levels. Aside from its improved performance, our method has two other advantages. (i) Simplicity, which implies that no additional hyperparameters are introduced. (ii) Our method is modular: it does not work in an end-to-end fashion, and can therefore be used to clean a dataset for any other future usage.

Recent work suggests that convolutional neural networks of different architectures learn to classify images in the same order. To understand this phenomenon, we revisit the over-parametrized deep linear network model. Our analysis reveals that, when the hidden layers are wide enough, the convergence rate of this model's parameters is exponentially faster along the directions of the larger principal components of the data, at a rate governed by the corresponding singular values. We term this convergence pattern the Principal Components bias (PC-bias). Empirically, we show how the PC-bias streamlines the order of learning of both linear and non-linear networks, more prominently at earlier stages of learning. We then compare our results to the simplicity bias, showing that both biases can be seen independently, and affect the order of learning in different ways. Finally, we discuss how the PC-bias may explain some benefits of early stopping and its connection to PCA, and why deep networks converge more slowly with random labels.

The learning trajectories of linguistic phenomena in humans provide insight into linguistic representation, beyond what can be gleaned from inspecting the behavior of an adult speaker. To apply a similar approach to analyze neural language models (NLM), it is first necessary to establish that different models are similar enough in the generalizations they make. In this paper, we show that NLMs with different initialization, architecture, and training data acquire linguistic phenomena in a similar order, despite their different end performance. These findings suggest that there is some mutual inductive bias that underlies these models' learning of linguistic phenomena. Taking inspiration from psycholinguistics, we argue that studying this inductive bias is an opportunity to study the linguistic representation implicit in NLMs. Leveraging these findings, we compare the relative performance on different phenomena at varying learning stages with simpler reference models. Results suggest that NLMs exhibit consistent "developmental" stages. Moreover, we find the learning trajectory to be approximately one-dimensional: given an NLM with a certain overall performance, it is possible to predict what linguistic generalizations it has already acquired. Initial analysis of these stages presents phenomena clusters (notably morphological ones), whose performance progresses in unison, suggesting a potential link between the generalizations behind them.

Deep active learning aims to reduce the annotation cost for the training of deep models, which is notoriously data-hungry. Until recently, deep active learning methods were ineffectual in the low-budget regime, where only a small number of examples are annotated. The situation has been alleviated by recent advances in representation and self-supervised learning, which impart the geometry of the data representation with rich information about the points. Taking advantage of this progress, we study the problem of subset selection for annotation through a "covering" lens, proposing ProbCover - a new active learning algorithm for the low budget regime, which seeks to maximize Probability Coverage. We then describe a dual way to view the proposed formulation, from which one can derive strategies suitable for the high budget regime of active learning, related to existing methods like Coreset. We conclude with extensive experiments, evaluating ProbCover in the low-budget regime. We show that our principled active learning strategy improves the state-of-the-art in the low-budget regime in several image recognition benchmarks. This method is especially beneficial in the semi-supervised setting, allowing state-of-the-art semi-supervised methods to match the performance of fully supervised methods, while using much fewer labels nonetheless. Code is available at https://github.com/avihu111/TypiClust.

Investigating active learning, we focus on the relation between the number of labeled examples (budget size), and suitable querying strategies. Our theoretical analysis shows a behavior reminiscent of phase transition: typical examples are best queried when the budget is low, while unrepresentative examples are best queried when the budget is large. Combined evidence shows that a similar phenomenon occurs in common classification models. Accordingly, we propose TypiClust -- a deep active learning strategy suited for low budgets. In a comparative empirical investigation of supervised learning, using a variety of architectures and image datasets, TypiClust outperforms all other active learning strategies in the low-budget regime. Using TypiClust in the semi-supervised framework, performance gets an even more significant boost. In particular, state-of-the-art semi-supervised methods trained on CIFAR-10 with 10 labeled examples selected by TypiClust, reach 93.2% accuracy -- an improvement of 39.4% over random selection. Code is available at https://github.com/avihu111/TypiClust.

Overfit is a fundamental problem in machine learning in general, and in deep learning in particular. In order to reduce overfit and improve generalization in the classification of images, some employ invariance to a group of transformations, such as rotations and reflections. However, since not all objects exhibit necessarily the same invariance, it seems desirable to allow the network to learn the useful level of invariance from the data. To this end, motivated by self-supervision, we introduce an architecture enhancement for existing neural network models based on input transformations, termed 'TransNet', together with a training algorithm suitable for it. Our model can be employed during training time only and then pruned for prediction, resulting in an equivalent architecture to the base model. Thus pruned, we show that our model improves performance on various data-sets while exhibiting improved generalization, which is achieved in turn by enforcing soft invariance on the convolutional kernels of the last layer in the base model. Theoretical analysis is provided to support the proposed method.