Christopher Kanan

Jayanta Dey, Shikhar Srivastava, Itamar Lerner et al.

Learning long-range non-stationary temporal patterns remains a core challenge for modern sequence models, particularly in strict streaming settings. In these settings, data arrive sequentially and must be processed in a single pass without simultaneously revisiting past observations. Standard architectures, including recurrent neural networks and transformers, are constrained by either truncated backpropagation through time horizon or explicit input window length for long range credit assignment. To address these limitations, we propose SHARP (Sleep-based Hierarchical Accelerated Replay), a framework that decomposes temporal learning into two complementary components: a memory module that accumulates a structured history of past inputs, and a pattern-recognition module that operates over this memory. This separation enables resource- and compute-efficient adaptation to non-stationary dynamics by eliminating the need for backpropagation through time across many steps for long-range credit assignment. Inspired by the accelerated replay observed in rodents during slow-wave sleep, SHARP incorporates offline (sleep) phases in which temporally structured memory traces are replayed in an accelerated form and integrated into higher-level memory representations, improving long-range context retention. Through controlled simulations and ablation studies, we characterize the key properties of the proposed framework. In benchmark datasets such as text8 and PG-19, we demonstrate that SHARP improves over recurrent baselines by retaining next-token predictive performance on previously seen data while continuing to learn from the current stream and generalizing to future unseen data. These gains are enabled by its hierarchical structure, which yields an exponentially increasing effective temporal context with only linear-time computational cost.

Eli Verwimp, Rahaf Aljundi, Shai Ben-David et al. · deepmind

Continual learning is a subfield of machine learning, which aims to allow machine learning models to continuously learn on new data, by accumulating knowledge without forgetting what was learned in the past. In this work, we take a step back, and ask: "Why should one care about continual learning in the first place?". We set the stage by examining recent continual learning papers published at four major machine learning conferences, and show that memory-constrained settings dominate the field. Then, we discuss five open problems in machine learning, and even though they might seem unrelated to continual learning at first sight, we show that continual learning will inevitably be part of their solution. These problems are model editing, personalization and specialization, on-device learning, faster (re-)training and reinforcement learning. Finally, by comparing the desiderata from these unsolved problems and the current assumptions in continual learning, we highlight and discuss four future directions for continual learning research. We hope that this work offers an interesting perspective on the future of continual learning, while displaying its potential value and the paths we have to pursue in order to make it successful. This work is the result of the many discussions the authors had at the Dagstuhl seminar on Deep Continual Learning, in March 2023.

Eugene Vorontsov, Alican Bozkurt, Adam Casson et al.

The use of artificial intelligence to enable precision medicine and decision support systems through the analysis of pathology images has the potential to revolutionize the diagnosis and treatment of cancer. Such applications will depend on models' abilities to capture the diverse patterns observed in pathology images. To address this challenge, we present Virchow, a foundation model for computational pathology. Using self-supervised learning empowered by the DINOv2 algorithm, Virchow is a vision transformer model with 632 million parameters trained on 1.5 million hematoxylin and eosin stained whole slide images from diverse tissue and specimen types, which is orders of magnitude more data than previous works. The Virchow model enables the development of a pan-cancer detection system with 0.949 overall specimen-level AUC across 17 different cancer types, while also achieving 0.937 AUC on 7 rare cancer types. The Virchow model sets the state-of-the-art on the internal and external image tile level benchmarks and slide level biomarker prediction tasks. The gains in performance highlight the importance of training on massive pathology image datasets, suggesting scaling up the data and network architecture can improve the accuracy for many high-impact computational pathology applications where limited amounts of training data are available.

Vaggelis Dorovatas, Malte Schwerin, Andrew D. Bagdanov et al. · mila

Foundation models have transformed machine learning through large-scale pretraining and increased test-time compute. Despite surpassing human performance in several domains, these models remain fundamentally limited in continuous operation, experience accumulation, and personalization, capabilities that are central to adaptive intelligence. While continual learning research has long targeted these goals, its historical focus on in-weight learning (IWL), i.e., updating a single model's parameters to absorb new knowledge, has rendered catastrophic forgetting a persistent challenge. Our position is that combining the strengths of In-Weight Learning (IWL) and the newly emerged capabilities of In-Context Learning (ICL) through the design of modular memory is the missing piece for continual adaptation at scale. We outline a conceptual framework for modular memory-centric architectures that leverage ICL for rapid adaptation and knowledge accumulation, and IWL for stable updates to model capabilities, charting a practical roadmap toward continually learning agents.

Tyler L. Hayes, Christopher Kanan

Real-time on-device continual learning is needed for new applications such as home robots, user personalization on smartphones, and augmented/virtual reality headsets. However, this setting poses unique challenges: embedded devices have limited memory and compute capacity and conventional machine learning models suffer from catastrophic forgetting when updated on non-stationary data streams. While several online continual learning models have been developed, their effectiveness for embedded applications has not been rigorously studied. In this paper, we first identify criteria that online continual learners must meet to effectively perform real-time, on-device learning. We then study the efficacy of several online continual learning methods when used with mobile neural networks. We measure their performance, memory usage, compute requirements, and ability to generalize to out-of-domain inputs.

Aneesh Rangnekar, Christopher Kanan, Matthew Hoffman

Using deep learning, we now have the ability to create exceptionally good semantic segmentation systems; however, collecting the prerequisite pixel-wise annotations for training images remains expensive and time-consuming. Therefore, it would be ideal to minimize the number of human annotations needed when creating a new dataset. Here, we address this problem by proposing a novel algorithm that combines active learning and semi-supervised learning. Active learning is an approach for identifying the best unlabeled samples to annotate. While there has been work on active learning for segmentation, most methods require annotating all pixel objects in each image, rather than only the most informative regions. We argue that this is inefficient. Instead, our active learning approach aims to minimize the number of annotations per-image. Our method is enriched with semi-supervised learning, where we use pseudo labels generated with a teacher-student framework to identify image regions that help disambiguate confused classes. We also integrate mechanisms that enable better performance on imbalanced label distributions, which have not been studied previously for active learning in semantic segmentation. In experiments on the CamVid and CityScapes datasets, our method obtains over 95% of the network's performance on the full-training set using less than 17% of the training data, whereas the previous state of the art required 40% of the training data.

Md Yousuf Harun, Jhair Gallardo, Tyler L. Hayes et al.

Supervised Continual learning involves updating a deep neural network (DNN) from an ever-growing stream of labeled data. While most work has focused on overcoming catastrophic forgetting, one of the major motivations behind continual learning is being able to efficiently update a network with new information, rather than retraining from scratch on the training dataset as it grows over time. Despite recent continual learning methods largely solving the catastrophic forgetting problem, there has been little attention paid to the efficiency of these algorithms. Here, we study recent methods for incremental class learning and illustrate that many are highly inefficient in terms of compute, memory, and storage. Some methods even require more compute than training from scratch! We argue that for continual learning to have real-world applicability, the research community cannot ignore the resources used by these algorithms. There is more to continual learning than mitigating catastrophic forgetting.

Md Yousuf Harun, Jhair Gallardo, Tyler L. Hayes et al.

In supervised continual learning, a deep neural network (DNN) is updated with an ever-growing data stream. Unlike the offline setting where data is shuffled, we cannot make any distributional assumptions about the data stream. Ideally, only one pass through the dataset is needed for computational efficiency. However, existing methods are inadequate and make many assumptions that cannot be made for real-world applications, while simultaneously failing to improve computational efficiency. In this paper, we propose a novel continual learning method, SIESTA based on wake/sleep framework for training, which is well aligned to the needs of on-device learning. The major goal of SIESTA is to advance compute efficient continual learning so that DNNs can be updated efficiently using far less time and energy. The principal innovations of SIESTA are: 1) rapid online updates using a rehearsal-free, backpropagation-free, and data-driven network update rule during its wake phase, and 2) expedited memory consolidation using a compute-restricted rehearsal policy during its sleep phase. For memory efficiency, SIESTA adapts latent rehearsal using memory indexing from REMIND. Compared to REMIND and prior arts, SIESTA is far more computationally efficient, enabling continual learning on ImageNet-1K in under 2 hours on a single GPU; moreover, in the augmentation-free setting it matches the performance of the offline learner, a milestone critical to driving adoption of continual learning in real-world applications.

Robik Shrestha, Kushal Kafle, Christopher Kanan

Dataset bias and spurious correlations can significantly impair generalization in deep neural networks. Many prior efforts have addressed this problem using either alternative loss functions or sampling strategies that focus on rare patterns. We propose a new direction: modifying the network architecture to impose inductive biases that make the network robust to dataset bias. Specifically, we propose OccamNets, which are biased to favor simpler solutions by design. OccamNets have two inductive biases. First, they are biased to use as little network depth as needed for an individual example. Second, they are biased toward using fewer image locations for prediction. While OccamNets are biased toward simpler hypotheses, they can learn more complex hypotheses if necessary. In experiments, OccamNets outperform or rival state-of-the-art methods run on architectures that do not incorporate these inductive biases. Furthermore, we demonstrate that when the state-of-the-art debiasing methods are combined with OccamNets results further improve.

Md Yousuf Harun, Jhair Gallardo, Junyu Chen et al.

Continual learning (CL) in deep neural networks (DNNs) involves incrementally accumulating knowledge in a DNN from a growing data stream. A major challenge in CL is that non-stationary data streams cause catastrophic forgetting of previously learned abilities. A popular solution is rehearsal: storing past observations in a buffer and then sampling the buffer to update the DNN. Uniform sampling in a class-balanced manner is highly effective, and better sample selection policies have been elusive. Here, we propose a new sample selection policy called GRASP that selects the most prototypical (easy) samples first and then gradually selects less prototypical (harder) examples. GRASP has little additional compute or memory overhead compared to uniform selection, enabling it to scale to large datasets. Compared to 17 other rehearsal policies, GRASP achieves higher accuracy in CL experiments on ImageNet. Compared to uniform balanced sampling, GRASP achieves the same performance with 40% fewer updates. We also show that GRASP is effective for CL on five text classification datasets.

Md Yousuf Harun, Christopher Kanan

Pre-trained deep neural networks (DNNs) are being widely deployed by industry for making business decisions and to serve users; however, a major problem is model decay, where the DNN's predictions become more erroneous over time, resulting in revenue loss or unhappy users. To mitigate model decay, DNNs are retrained from scratch using old and new data. This is computationally expensive, so retraining happens only once performance significantly decreases. Here, we study how continual learning (CL) could potentially overcome model decay in large pre-trained DNNs and greatly reduce computational costs for keeping DNNs up-to-date. We identify the "stability gap" as a major obstacle in our setting. The stability gap refers to a phenomenon where learning new data causes large drops in performance for past tasks before CL mitigation methods eventually compensate for this drop. We test two hypotheses to investigate the factors influencing the stability gap and identify a method that vastly reduces this gap. In large-scale experiments for both easy and hard CL distributions (e.g., class incremental learning), we demonstrate that our method reduces the stability gap and greatly increases computational efficiency. Our work aligns CL with the goals of the production setting, where CL is needed for many applications.

Rakshit S. Kothari, Reynold J. Bailey, Christopher Kanan et al.

The study of human gaze behavior in natural contexts requires algorithms for gaze estimation that are robust to a wide range of imaging conditions. However, algorithms often fail to identify features such as the iris and pupil centroid in the presence of reflective artifacts and occlusions. Previous work has shown that convolutional networks excel at extracting gaze features despite the presence of such artifacts. However, these networks often perform poorly on data unseen during training. This work follows the intuition that jointly training a convolutional network with multiple datasets learns a generalized representation of eye parts. We compare the performance of a single model trained with multiple datasets against a pool of models trained on individual datasets. Results indicate that models tested on datasets in which eye images exhibit higher appearance variability benefit from multiset training. In contrast, dataset-specific models generalize better onto eye images with lower appearance variability.

Indranil Sur, Zachary Daniels, Abrar Rahman et al.

As Artificial and Robotic Systems are increasingly deployed and relied upon for real-world applications, it is important that they exhibit the ability to continually learn and adapt in dynamically-changing environments, becoming Lifelong Learning Machines. Continual/lifelong learning (LL) involves minimizing catastrophic forgetting of old tasks while maximizing a model's capability to learn new tasks. This paper addresses the challenging lifelong reinforcement learning (L2RL) setting. Pushing the state-of-the-art forward in L2RL and making L2RL useful for practical applications requires more than developing individual L2RL algorithms; it requires making progress at the systems-level, especially research into the non-trivial problem of how to integrate multiple L2RL algorithms into a common framework. In this paper, we introduce the Lifelong Reinforcement Learning Components Framework (L2RLCF), which standardizes L2RL systems and assimilates different continual learning components (each addressing different aspects of the lifelong learning problem) into a unified system. As an instantiation of L2RLCF, we develop a standard API allowing easy integration of novel lifelong learning components. We describe a case study that demonstrates how multiple independently-developed LL components can be integrated into a single realized system. We also introduce an evaluation environment in order to measure the effect of combining various system components. Our evaluation environment employs different LL scenarios (sequences of tasks) consisting of Starcraft-2 minigames and allows for the fair, comprehensive, and quantitative comparison of different combinations of components within a challenging common evaluation environment.

Tyler L. Hayes, Maximilian Nickel, Christopher Kanan et al.

There has been significant progress in creating machine learning models that identify objects in scenes along with their associated attributes and relationships; however, there is a large gap between the best models and human capabilities. One of the major reasons for this gap is the difficulty in collecting sufficient amounts of annotated relations and attributes for training these systems. While some attributes and relations are abundant, the distribution in the natural world and existing datasets is long tailed. In this paper, we address this problem by introducing a novel incremental active learning framework that asks for attributes and relations in visual scenes. While conventional active learning methods ask for labels of specific examples, we flip this framing to allow agents to ask for examples from specific categories. Using this framing, we introduce an active sampling method that asks for examples from the tail of the data distribution and show that it outperforms classical active learning methods on Visual Genome.

Aneesh Rangnekar, Christopher Kanan, Matthew Hoffman

Obtaining human per-pixel labels for semantic segmentation is incredibly laborious, often making labeled dataset construction prohibitively expensive. Here, we endeavor to overcome this problem with a novel algorithm that combines semi-supervised and active learning, resulting in the ability to train an effective semantic segmentation algorithm with significantly lesser labeled data. To do this, we extend the prior state-of-the-art S4AL algorithm by replacing its mean teacher approach for semi-supervised learning with a self-training approach that improves learning with noisy labels. We further boost the neural network's ability to query useful data by adding a contrastive learning head, which leads to better understanding of the objects in the scene, and hence, better queries for active learning. We evaluate our method on CamVid and CityScapes datasets, the de-facto standards for active learning for semantic segmentation. We achieve more than 95% of the network's performance on CamVid and CityScapes datasets, utilizing only 12.1% and 15.1% of the labeled data, respectively. We also benchmark our method across existing stand-alone semi-supervised learning methods on the CityScapes dataset and achieve superior performance without any bells or whistles.

Matthew Iceland, Christopher Kanan

Image Augmentations are widely used to reduce overfitting in neural networks. However, the explainability of their benefits largely remains a mystery. We study which layers of residual neural networks (ResNets) are most affected by augmentations using Centered Kernel Alignment (CKA). We do so by analyzing models of varying widths and depths, as well as whether their weights are initialized randomly or through transfer learning. We find that the pattern of how the layers are affected depends on the model's depth, and that networks trained with augmentation that use information from two images affect the learned weights significantly more than augmentations that operate on a single image. Deeper layers of ResNets initialized with ImageNet-1K weights and fine-tuned receive more impact from the augmentations than early layers. Understanding the effects of image augmentations on CNNs will have a variety of applications, such as determining how far back one needs to fine-tune a network and which layers should be frozen when implementing layer freezing algorithms.

Junyu Chen, Jie An, Hanjia Lyu et al.

Visual-textual sentiment analysis aims to predict sentiment with the input of a pair of image and text, which poses a challenge in learning effective features for diverse input images. To address this, we propose a holistic method that achieves robust visual-textual sentiment analysis by exploiting a rich set of powerful pre-trained visual and textual prior models. The proposed method consists of four parts: (1) a visual-textual branch to learn features directly from data for sentiment analysis, (2) a visual expert branch with a set of pre-trained "expert" encoders to extract selected semantic visual features, (3) a CLIP branch to implicitly model visual-textual correspondence, and (4) a multimodal feature fusion network based on BERT to fuse multimodal features and make sentiment predictions. Extensive experiments on three datasets show that our method produces better visual-textual sentiment analysis performance than existing methods.

Jian Lu, Shikhar Srivastava, Junyu Chen et al.

With the advent of multi-modal large language models (MLLMs), datasets used for visual question answering (VQA) and referring expression comprehension have seen a resurgence. However, the most popular datasets used to evaluate MLLMs are some of the earliest ones created, and they have many known problems, including extreme bias, spurious correlations, and an inability to permit fine-grained analysis. In this paper, we pioneer evaluating recent MLLMs (LLaVA 1.5, LLaVA-NeXT, BLIP2, InstructBLIP, GPT-4V, and GPT-4o) on datasets designed to address weaknesses in earlier ones. We assess three VQA datasets: 1) TDIUC, which permits fine-grained analysis on 12 question types; 2) TallyQA, which has simple and complex counting questions; and 3) DVQA, which requires optical character recognition for chart understanding. We also study VQDv1, a dataset that requires identifying all image regions that satisfy a given query. Our experiments reveal the weaknesses of many MLLMs that have not previously been reported. Our code is integrated into the widely used LAVIS framework for MLLM evaluation, enabling the rapid assessment of future MLLMs. Project webpage: https://kevinlujian.github.io/MLLM_Evaluations/

Rahman Ejaz, Varchas Gopalaswamy, Ricardo Luna et al.

Bayesian optimization (BO) has for sequential optimization of expensive black-box functions demonstrated practicality and effectiveness in many real-world settings. Meta-Bayesian optimization (meta-BO) focuses on improving the sample efficiency of BO by making use of information from related tasks. Although meta-BO is sample-efficient when task structure transfers, poor alignment between meta-training and test tasks can cause suboptimal queries to be suggested during online optimization. To this end, we propose a simple meta-BO algorithm that utilizes related-task information when determined useful, falling back to lookahead otherwise, within a unified framework. We demonstrate competitiveness of our method with existing approaches on function optimization tasks, while retaining strong performance in low task-relatedness regimes where test tasks share limited structure with the meta-training set.

Yi Kan Wang, Ludmila Tydlitatova, Jeremy D. Kunz et al.

Molecular assays are standard of care for detecting genomic alterations in cancer prognosis and therapy selection but are costly, tissue-destructive and time-consuming. Artificial intelligence (AI) applied to routine hematoxylin and eosin (H&E)-stained whole slide images (WSIs) offers a fast and economical alternative for screening molecular biomarkers. We introduce OmniScreen, a high-throughput AI-based system leveraging Virchow2 embeddings extracted from 60,529 cancer patients with paired 489-gene MSK-IMPACT targeted biomarker panel and WSIs. Unlike conventional approaches that train separate models for each biomarker, OmniScreen employs a unified model to predict a broad range of clinically relevant biomarkers across cancers, including low-prevalence targets impractical to model individually. OmniScreen reliably identifies therapeutic targets and shared phenotypic features across common and rare tumors. We investigate the biomarker prediction probabilities and accuracies of OmniScreen in relation to tumor area, cohort size, histologic subtype alignment, and pathway-level morphological patterns. These findings underscore the potential of OmniScreen for routine clinical screening.

Rahman Ejaz, Varchas Gopalaswamy, Riccardo Betti et al.

The domain of laser fusion presents a unique and challenging predictive modeling application landscape for machine learning methods due to high problem complexity and limited training data. Data-driven approaches utilizing prescribed functional forms, inductive biases and physics-informed learning (PIL) schemes have been successful in the past for achieving desired generalization ability and model interpretation that aligns with physics expectations. In complex multi-physics application domains, however, it is not always obvious how architectural biases or discriminative penalties can be formulated. In this work, focusing on nuclear fusion energy using high powered lasers, we present the use of Kolmogorov-Arnold Networks (KANs) as an alternative to PIL for developing a new type of data-driven predictive model which is able to achieve high prediction accuracy and physics interpretability. A KAN based model, a MLP with PIL, and a baseline MLP model are compared in generalization ability and interpretation with a domain expert-derived symbolic regression model. Through empirical studies in this high physics complexity domain, we show that KANs can potentially provide benefits when developing predictive models for data-starved physics applications.

Abhishek Tyagi, Arjun Iyer, William H Renninger et al.

Recent advances in Dynamic Sparse Training (DST) have pushed the frontier of sparse neural network training in structured and unstructured contexts, matching dense-model performance while drastically reducing parameter counts to facilitate model scaling. However, unstructured sparsity often fails to translate into practical speedups on modern hardware. To address this shortcoming, we propose DynaDiag, a novel structured sparse-to-sparse DST method that performs at par with unstructured sparsity. DynaDiag enforces a diagonal sparsity pattern throughout training and preserves sparse computation in forward and backward passes. We further leverage the diagonal structure to accelerate computation via a custom CUDA kernel, rendering the method hardware-friendly. Empirical evaluations on diverse neural architectures demonstrate that our method maintains accuracy on par with unstructured counterparts while benefiting from tangible computational gains. Notably, with 90% sparse linear layers in ViTs, we observe up to a 3.13x speedup in online inference without sacrificing model performance and a 1.59x speedup in training on a GPU compared to equivalent unstructured layers. Our source code is available at https://github.com/horizon-research/DynaDiag/.

Md Saiful Islam, Ekram Hossain, Abdelrahman Abdelkader et al.

Video-based assessments offer a scalable pathway for remote Parkinson's disease (PD) screening. While traditional approaches rely on handcrafted features mimicking clinical scales, recent advances in video foundation models (VFMs) enable representation learning without task-specific customization. However, the comparative effectiveness of different VFM architectures across diverse clinical tasks remains poorly understood. We present a large-scale systematic study using a novel video dataset from 1,888 participants (727 with PD), comprising 32,847 videos across 16 standardized clinical tasks. We evaluate seven state-of-the-art VFMs -- including VideoPrism, V-JEPA, ViViT, and VideoMAE -- to determine their robustness in clinical screening. By evaluating frozen embeddings with a linear classification head, we demonstrate that task saliency is highly model-dependent: VideoPrism excels in capturing visual speech kinematics (no audio) and facial expressivity, while V-JEPA proves superior for upper-limb motor tasks. Notably, TimeSformer remains highly competitive for rhythmic tasks like finger tapping. Our experiments yield AUCs of 76.4 - 85.3% and accuracies of 71.5 - 80.6%. While high specificity (up to 90.3%) suggests strong potential for ruling out healthy individuals, the lower sensitivity (43.2 - 57.3%) highlights the need for task-aware calibration and integration of multiple tasks and modalities. Overall, this work establishes a rigorous baseline for VFM-based PD screening and provides a roadmap for selecting suitable tasks and architectures in remote neurological monitoring. Code and anonymized structured data are publicly available: https://anonymous.4open.science/r/parkinson\_video\_benchmarking-A2C5

Katarzyna Filus, Kamil Faber, Roberto Corizzo et al.

Continual learning studies how models can adapt to new tasks while retaining previously acquired knowledge. Although a broad spectrum of methods has been proposed to mitigate catastrophic forgetting, the field remains predominantly performance-driven, with limited insight into what forgetting actually corresponds to within the vision model's representation space. Prior work has primarily analyzed forgetting through task-level performance or coarse measures of representational drift, without disentangling output-level accessibility from changes in finer-grained internal structure. To this end, we propose a diagnostic framework that leverages Sparse Autoencoders (SAEs) to define a task-anchored latent feature space, enabling analysis of how task-specific information evolves at a finer granularity, where individual SAE latents are treated as concept proxies for recurring and relatively disentangled visual patterns in the model's internal computations. Within this framework, we decompose forgetting into apparent concept deletion, recoverability, and decodability. We show that a large portion of seemingly lost concept-level information can often be recovered under linearity assumption, with concept decodability degrading as more tasks are introduced. Overall, our findings suggest that a significant part of concept-level forgetting can be attributed to changes in the representational accessibility rather than complete information erasure.

Vincenzo Lomonaco, Lorenzo Pellegrini, Andrea Cossu et al.

Learning continually from non-stationary data streams is a long-standing goal and a challenging problem in machine learning. Recently, we have witnessed a renewed and fast-growing interest in continual learning, especially within the deep learning community. However, algorithmic solutions are often difficult to re-implement, evaluate and port across different settings, where even results on standard benchmarks are hard to reproduce. In this work, we propose Avalanche, an open-source end-to-end library for continual learning research based on PyTorch. Avalanche is designed to provide a shared and collaborative codebase for fast prototyping, training, and reproducible evaluation of continual learning algorithms.

Robik Shrestha, Kushal Kafle, Christopher Kanan

A critical problem in deep learning is that systems learn inappropriate biases, resulting in their inability to perform well on minority groups. This has led to the creation of multiple algorithms that endeavor to mitigate bias. However, it is not clear how effective these methods are. This is because study protocols differ among papers, systems are tested on datasets that fail to test many forms of bias, and systems have access to hidden knowledge or are tuned specifically to the test set. To address this, we introduce an improved evaluation protocol, sensible metrics, and a new dataset, which enables us to ask and answer critical questions about bias mitigation algorithms. We evaluate seven state-of-the-art algorithms using the same network architecture and hyperparameter selection policy across three benchmark datasets. We introduce a new dataset called Biased MNIST that enables assessment of robustness to multiple bias sources. We use Biased MNIST and a visual question answering (VQA) benchmark to assess robustness to hidden biases. Rather than only tuning to the test set distribution, we study robustness across different tuning distributions, which is critical because for many applications the test distribution may not be known during development. We find that algorithms exploit hidden biases, are unable to scale to multiple forms of bias, and are highly sensitive to the choice of tuning set. Based on our findings, we implore the community to adopt more rigorous assessment of future bias mitigation methods. All data, code, and results are publicly available at: https://github.com/erobic/bias-mitigators.

Aneesh Rangnekar, Nilay Mokashi, Emmett Ientilucci et al.

We investigate applying convolutional neural network (CNN) architecture to facilitate aerial hyperspectral scene understanding and present a new hyperspectral dataset-AeroRIT-that is large enough for CNN training. To date the majority of hyperspectral airborne have been confined to various sub-categories of vegetation and roads and this scene introduces two new categories: buildings and cars. To the best of our knowledge, this is the first comprehensive large-scale hyperspectral scene with nearly seven million pixel annotations for identifying cars, roads, and buildings. We compare the performance of three popular architectures - SegNet, U-Net, and Res-U-Net, for scene understanding and object identification via the task of dense semantic segmentation to establish a benchmark for the scene. To further strengthen the network, we add squeeze and excitation blocks for better channel interactions and use self-supervised learning for better encoder initialization. Aerial hyperspectral image analysis has been restricted to small datasets with limited train/test splits capabilities and we believe that AeroRIT will help advance the research in the field with a more complex object distribution to perform well on. The full dataset, with flight lines in radiance and reflectance domain, is available for download at https://github.com/aneesh3108/AeroRIT. This dataset is the first step towards developing robust algorithms for hyperspectral airborne sensing that can robustly perform advanced tasks like vehicle tracking and occlusion handling.

George Shaikovski, Adam Casson, Kristen Severson et al.

Foundation models in computational pathology promise to unlock the development of new clinical decision support systems and models for precision medicine. However, there is a mismatch between most clinical analysis, which is defined at the level of one or more whole slide images, and foundation models to date, which process the thousands of image tiles contained in a whole slide image separately. The requirement to train a network to aggregate information across a large number of tiles in multiple whole slide images limits these models' impact. In this work, we present a slide-level foundation model for H&E-stained histopathology, PRISM, that builds on Virchow tile embeddings and leverages clinical report text for pre-training. Using the tile embeddings, PRISM produces slide-level embeddings with the ability to generate clinical reports, resulting in several modes of use. Using text prompts, PRISM achieves zero-shot cancer detection and sub-typing performance approaching and surpassing that of a supervised aggregator model. Using the slide embeddings with linear classifiers, PRISM surpasses supervised aggregator models. Furthermore, we demonstrate that fine-tuning of the PRISM slide encoder yields label-efficient training for biomarker prediction, a task that typically suffers from low availability of training data; an aggregator initialized with PRISM and trained on as little as 10% of the training data can outperform a supervised baseline that uses all of the data.

Md Yousuf Harun, Kyungbok Lee, Jhair Gallardo et al.

Embeddings produced by pre-trained deep neural networks (DNNs) are widely used; however, their efficacy for downstream tasks can vary widely. We study the factors influencing transferability and out-of-distribution (OOD) generalization of pre-trained DNN embeddings through the lens of the tunnel effect hypothesis, which is closely related to intermediate neural collapse. This hypothesis suggests that deeper DNN layers compress representations and hinder OOD generalization. Contrary to earlier work, our experiments show this is not a universal phenomenon. We comprehensively investigate the impact of DNN architecture, training data, image resolution, and augmentations on transferability. We identify that training with high-resolution datasets containing many classes greatly reduces representation compression and improves transferability. Our results emphasize the danger of generalizing findings from toy datasets to broader contexts.

Md Yousuf Harun, Jhair Gallardo, Christopher Kanan

Out-of-distribution (OOD) detection and OOD generalization are widely studied in Deep Neural Networks (DNNs), yet their relationship remains poorly understood. We empirically show that the degree of Neural Collapse (NC) in a network layer is inversely related with these objectives: stronger NC improves OOD detection but degrades generalization, while weaker NC enhances generalization at the cost of detection. This trade-off suggests that a single feature space cannot simultaneously achieve both tasks. To address this, we develop a theoretical framework linking NC to OOD detection and generalization. We show that entropy regularization mitigates NC to improve generalization, while a fixed Simplex Equiangular Tight Frame (ETF) projector enforces NC for better detection. Based on these insights, we propose a method to control NC at different DNN layers. In experiments, our method excels at both tasks across OOD datasets and DNN architectures. Code for our experiments is available at: https://yousuf907.github.io/ncoodg

Md Yousuf Harun, Christopher Kanan

To adapt to real-world data streams, continual learning (CL) systems must rapidly learn new concepts while preserving and utilizing prior knowledge. When it comes to adding new information to continually-trained deep neural networks (DNNs), classifier weights for newly encountered categories are typically initialized randomly, leading to high initial training loss (spikes) and instability. Consequently, achieving optimal convergence and accuracy requires prolonged training, increasing computational costs. Inspired by Neural Collapse (NC), we propose a weight initialization strategy to improve learning efficiency in CL. In DNNs trained with mean-squared-error, NC gives rise to a Least-Square (LS) classifier in the last layer, whose weights can be analytically derived from learned features. We leverage this LS formulation to initialize classifier weights in a data-driven manner, aligning them with the feature distribution rather than using random initialization. Our method mitigates initial loss spikes and accelerates adaptation to new tasks. We evaluate our approach in large-scale CL settings, demonstrating faster adaptation and improved CL performance.

Shikhar Srivastava, Md Yousuf Harun, Robik Shrestha et al.

Generative large language models (LLMs) exhibit impressive capabilities, which can be further augmented by integrating a pre-trained vision model into the original LLM to create a multimodal LLM (MLLM). However, this integration often significantly decreases performance on natural language understanding and generation tasks, compared to the original LLM. This study investigates this issue using the LLaVA MLLM, treating the integration as a continual learning problem. We evaluate five continual learning methods to mitigate forgetting and identify a technique that enhances visual understanding while minimizing linguistic performance loss. Our approach reduces linguistic performance degradation by up to 15% over the LLaVA recipe, while maintaining high multimodal accuracy. We also demonstrate the robustness of our method through continual learning on a sequence of vision-language tasks, effectively preserving linguistic skills while acquiring new multimodal capabilities. Project webpage: https://shikhar-srivastava.github.io/cl-for-improving-mllms

Abhishek Tyagi, Arjun Iyer, Liam Young et al.

Structured sparsity accelerates training and inference on modern GPUs, yet it still trails unstructured dynamic sparse training (DST) in accuracy. The shortfall stems from a loss of expressivity: whereas a dense layer can realize every possible mask obtained by choosing any $w$ active weights out of $n$, a fixed block or N:M layout explores only a subset of those possibilities. We propose to close this gap by learning, for each layer, a single permutation matrix jointly with the structured weight matrix. Applied to three canonical structures -- block, N:M, and diagonals -- we show that permutation-augmented DST (PA-DST) matches unstructured baselines (RigL, SET) at 90--95\% sparsity on ImageNet-1K (ViT-B/16) and WikiText-103 (GPT-2), yet trains up to $1.21\times$ and infers up to $2.9\times$ faster. The results position structure + learned permutation as a sweet spot between accuracy and efficiency.

Jayanta Dey, Nicholas Soures, Miranda Gonzales et al.

In this pilot study, we propose a neuro-inspired approach that compresses temporal sequences into context-tagged chunks, where each tag represents a recurring structural unit or``community'' in the sequence. These tags are generated during an offline sleep phase and serve as compact references to past experience, allowing the learner to incorporate information beyond its immediate input range. We evaluate this idea in a controlled synthetic environment designed to reveal the limitations of traditional neural network based sequence learners, such as recurrent neural networks (RNNs), when facing temporal patterns on multiple timescales. We evaluate this idea in a controlled synthetic environment designed to reveal the limitations of traditional neural network based sequence learners, such as recurrent neural networks (RNNs), when facing temporal patterns on multiple timescales. Our results, while preliminary, suggest that temporal chunking can significantly enhance learning efficiency under resource constrained settings. A small-scale human pilot study using a Serial Reaction Time Task further motivates the idea of structural abstraction. Although limited to synthetic tasks, this work serves as an early proof-of-concept, with initial evidence that learned context tags can transfer across related task, offering potential for future applications in transfer learning.

Wenbo Zhang, Junyu Chen, Christopher Kanan

Due to their large sizes, volumetric scans and whole-slide pathology images (WSIs) are often processed by extracting embeddings from local regions and then an aggregator makes predictions from this set. However, current methods require post-hoc visualization techniques (e.g., Grad-CAM) and often fail to localize small yet clinically crucial details. To address these limitations, we introduce INSIGHT, a novel weakly-supervised aggregator that integrates heatmap generation as an inductive bias. Starting from pre-trained feature maps, INSIGHT employs a detection module with small convolutional kernels to capture fine details and a context module with a broader receptive field to suppress local false positives. The resulting internal heatmap highlights diagnostically relevant regions. On CT and WSI benchmarks, INSIGHT achieves state-of-the-art classification results and high weakly-labeled semantic segmentation performance. Project website and code are available at: https://zhangdylan83.github.io/ewsmia/

Chen Ding, Christopher Kanan, Dylan McKellips et al.

The greatest demand for today's computing is machine learning. This paper analyzes three machine learning algorithms: transformers, spatial convolution, and FFT. The analysis is novel in three aspects. First, it measures the cost of memory access on an abstract memory hierarchy, instead of traditional time or space complexity. Second, the analysis is asymptotic and identifies the primary sources of the memory cost. Finally, the result is symbolic, which can be used to select algorithmic parameters such as the group size in grouped query attention for any dimension size and number of heads and the batch size for batched convolution for any image size and kernel size.

Yunye Gong, Robik Shrestha, Jared Claypoole et al.

We propose a novel VQA dataset, BloomVQA, to facilitate comprehensive evaluation of large vision-language models on comprehension tasks. Unlike current benchmarks that often focus on fact-based memorization and simple reasoning tasks without theoretical grounding, we collect multiple-choice samples based on picture stories that reflect different levels of comprehension, as laid out in Bloom's Taxonomy, a classic framework for learning assessment widely adopted in education research. Our data maps to a novel hierarchical graph representation which enables automatic data augmentation and novel measures characterizing model consistency. We perform graded evaluation and reliability analysis on recent multi-modal models. In comparison to low-level tasks, we observe decreased performance on tasks requiring advanced comprehension and cognitive skills with up to 38.0\% drop in VQA accuracy. In comparison to earlier models, GPT-4V demonstrates improved accuracy over all comprehension levels and shows a tendency of bypassing visual inputs especially for higher-level tasks. Current models also show consistency patterns misaligned with human comprehension in various scenarios, demonstrating the need for improvement based on theoretically-grounded criteria.

Junyu Chen, Jie An, Hanjia Lyu et al.

Assessing the artness of AI-generated images continues to be a challenge within the realm of image generation. Most existing metrics cannot be used to perform instance-level and reference-free artness evaluation. This paper presents ArtScore, a metric designed to evaluate the degree to which an image resembles authentic artworks by artists (or conversely photographs), thereby offering a novel approach to artness assessment. We first blend pre-trained models for photo and artwork generation, resulting in a series of mixed models. Subsequently, we utilize these mixed models to generate images exhibiting varying degrees of artness with pseudo-annotations. Each photorealistic image has a corresponding artistic counterpart and a series of interpolated images that range from realistic to artistic. This dataset is then employed to train a neural network that learns to estimate quantized artness levels of arbitrary images. Extensive experiments reveal that the artness levels predicted by ArtScore align more closely with human artistic evaluation than existing evaluation metrics, such as Gram loss and ArtFID.

Manoj Acharya, Anirban Roy, Kaushik Koneripalli et al.

This paper presents an approach to detect out-of-context (OOC) objects in an image. Given an image with a set of objects, our goal is to determine if an object is inconsistent with the scene context and detect the OOC object with a bounding box. In this work, we consider commonly explored contextual relations such as co-occurrence relations, the relative size of an object with respect to other objects, and the position of the object in the scene. We posit that contextual cues are useful to determine object labels for in-context objects and inconsistent context cues are detrimental to determining object labels for out-of-context objects. To realize this hypothesis, we propose a graph contextual reasoning network (GCRN) to detect OOC objects. GCRN consists of two separate graphs to predict object labels based on the contextual cues in the image: 1) a representation graph to learn object features based on the neighboring objects and 2) a context graph to explicitly capture contextual cues from the neighboring objects. GCRN explicitly captures the contextual cues to improve the detection of in-context objects and identify objects that violate contextual relations. In order to evaluate our approach, we create a large-scale dataset by adding OOC object instances to the COCO images. We also evaluate on recent OCD benchmark. Our results show that GCRN outperforms competitive baselines in detecting OOC objects and correctly detecting in-context objects.

Manoj Acharya, Christopher Kanan

In this technical report, we present our approaches for the continual object detection track of the SODA10M challenge. We adapt ResNet50-FPN as the baseline and try several improvements for the final submission model. We find that task-specific replay scheme, learning rate scheduling, model calibration, and using original image scale helps to improve performance for both large and small objects in images. Our team `hypertune28' secured the second position among 52 participants in the challenge. This work will be presented at the ICCV 2021 Workshop on Self-supervised Learning for Next-Generation Industry-level Autonomous Driving (SSLAD).

Yipeng Zhang, Tyler L. Hayes, Christopher Kanan

Humans are incredibly good at transferring knowledge from one domain to another, enabling rapid learning of new tasks. Likewise, transfer learning has enabled enormous success in many computer vision problems using pretraining. However, the benefits of transfer in multi-domain learning, where a network learns multiple tasks defined by different datasets, has not been adequately studied. Learning multiple domains could be beneficial, or these domains could interfere with each other given limited network capacity. Understanding how deep neural networks of varied capacity facilitate transfer across inputs from different distributions is a critical step towards open world learning. In this work, we decipher the conditions where interference and knowledge transfer occur in multi-domain learning. We propose new metrics disentangling interference and transfer, set up experimental protocols, and examine the roles of network capacity, task grouping, and dynamic loss weighting in reducing interference and facilitating transfer.

Bidur Khanal, Christopher Kanan

Incorrectly labeled examples, or label noise, is common in real-world computer vision datasets. While the impact of label noise on learning in deep neural networks has been studied in prior work, these studies have exclusively focused on homogeneous label noise, i.e., the degree of label noise is the same across all categories. However, in the real-world, label noise is often heterogeneous, with some categories being affected to a greater extent than others. Here, we address this gap in the literature. We hypothesized that heterogeneous label noise would only affect the classes that had label noise unless there was transfer from those classes to the classes without label noise. To test this hypothesis, we designed a series of computer vision studies using MNIST, CIFAR-10, CIFAR-100, and MS-COCO where we imposed heterogeneous label noise during the training of multi-class, multi-task, and multi-label systems. Our results provide evidence in support of our hypothesis: label noise only affects the class affected by it unless there is transfer.

Tyler L. Hayes, Giri P. Krishnan, Maxim Bazhenov et al.

Replay is the reactivation of one or more neural patterns, which are similar to the activation patterns experienced during past waking experiences. Replay was first observed in biological neural networks during sleep, and it is now thought to play a critical role in memory formation, retrieval, and consolidation. Replay-like mechanisms have been incorporated into deep artificial neural networks that learn over time to avoid catastrophic forgetting of previous knowledge. Replay algorithms have been successfully used in a wide range of deep learning methods within supervised, unsupervised, and reinforcement learning paradigms. In this paper, we provide the first comprehensive comparison between replay in the mammalian brain and replay in artificial neural networks. We identify multiple aspects of biological replay that are missing in deep learning systems and hypothesize how they could be utilized to improve artificial neural networks.

Jhair Gallardo, Tyler L. Hayes, Christopher Kanan

In continual learning, a system must incrementally learn from a non-stationary data stream without catastrophic forgetting. Recently, multiple methods have been devised for incrementally learning classes on large-scale image classification tasks, such as ImageNet. State-of-the-art continual learning methods use an initial supervised pre-training phase, in which the first 10% - 50% of the classes in a dataset are used to learn representations in an offline manner before continual learning of new classes begins. We hypothesize that self-supervised pre-training could yield features that generalize better than supervised learning, especially when the number of samples used for pre-training is small. We test this hypothesis using the self-supervised MoCo-V2, Barlow Twins, and SwAV algorithms. On ImageNet, we find that these methods outperform supervised pre-training considerably for online continual learning, and the gains are larger when fewer samples are available. Our findings are consistent across three online continual learning algorithms. Our best system achieves a 14.95% relative increase in top-1 accuracy on class incremental ImageNet over the prior state of the art for online continual learning.

Tyler L. Hayes, Christopher Kanan

In continual learning, a system learns from non-stationary data streams or batches without catastrophic forgetting. While this problem has been heavily studied in supervised image classification and reinforcement learning, continual learning in neural networks designed for abstract reasoning has not yet been studied. Here, we study continual learning of analogical reasoning. Analogical reasoning tests such as Raven's Progressive Matrices (RPMs) are commonly used to measure non-verbal abstract reasoning in humans, and recently offline neural networks for the RPM problem have been proposed. In this paper, we establish experimental baselines, protocols, and forward and backward transfer metrics to evaluate continual learners on RPMs. We employ experience replay to mitigate catastrophic forgetting. Prior work using replay for image classification tasks has found that selectively choosing the samples to replay offers little, if any, benefit over random selection. In contrast, we find that selective replay can significantly outperform random selection for the RPM task.

Usman Mahmood, Robik Shrestha, David D. B. Bates et al.

Artificial intelligence (AI) has been successful at solving numerous problems in machine perception. In radiology, AI systems are rapidly evolving and show progress in guiding treatment decisions, diagnosing, localizing disease on medical images, and improving radiologists' efficiency. A critical component to deploying AI in radiology is to gain confidence in a developed system's efficacy and safety. The current gold standard approach is to conduct an analytical validation of performance on a generalization dataset from one or more institutions, followed by a clinical validation study of the system's efficacy during deployment. Clinical validation studies are time-consuming, and best practices dictate limited re-use of analytical validation data, so it is ideal to know ahead of time if a system is likely to fail analytical or clinical validation. In this paper, we describe a series of sanity tests to identify when a system performs well on development data for the wrong reasons. We illustrate the sanity tests' value by designing a deep learning system to classify pancreatic cancer seen in computed tomography scans.

Ryne Roady, Tyler L. Hayes, Christopher Kanan

Supervised classification methods often assume that evaluation data is drawn from the same distribution as training data and that all classes are present for training. However, real-world classifiers must handle inputs that are far from the training distribution including samples from unknown classes. Open set robustness refers to the ability to properly label samples from previously unseen categories as novel and avoid high-confidence, incorrect predictions. Existing approaches have focused on either novel inference methods, unique training architectures, or supplementing the training data with additional background samples. Here, we propose a simple regularization technique easily applied to existing convolutional neural network architectures that improves open set robustness without a background dataset. Our method achieves state-of-the-art results on open set classification baselines and easily scales to large-scale open set classification problems.

Manoj Acharya, Tyler L. Hayes, Christopher Kanan

Humans can incrementally learn to do new visual detection tasks, which is a huge challenge for today's computer vision systems. Incrementally trained deep learning models lack backwards transfer to previously seen classes and suffer from a phenomenon known as $"catastrophic forgetting."$ In this paper, we pioneer online streaming learning for object detection, where an agent must learn examples one at a time with severe memory and computational constraints. In object detection, a system must output all bounding boxes for an image with the correct label. Unlike earlier work, the system described in this paper can learn this task in an online manner with new classes being introduced over time. We achieve this capability by using a novel memory replay mechanism that efficiently replays entire scenes. We achieve state-of-the-art results on both the PASCAL VOC 2007 and MS COCO datasets.

Damien Teney, Kushal Kafle, Robik Shrestha et al.

Out-of-distribution (OOD) testing is increasingly popular for evaluating a machine learning system's ability to generalize beyond the biases of a training set. OOD benchmarks are designed to present a different joint distribution of data and labels between training and test time. VQA-CP has become the standard OOD benchmark for visual question answering, but we discovered three troubling practices in its current use. First, most published methods rely on explicit knowledge of the construction of the OOD splits. They often rely on ``inverting'' the distribution of labels, e.g. answering mostly 'yes' when the common training answer is 'no'. Second, the OOD test set is used for model selection. Third, a model's in-domain performance is assessed after retraining it on in-domain splits (VQA v2) that exhibit a more balanced distribution of labels. These three practices defeat the objective of evaluating generalization, and put into question the value of methods specifically designed for this dataset. We show that embarrassingly-simple methods, including one that generates answers at random, surpass the state of the art on some question types. We provide short- and long-term solutions to avoid these pitfalls and realize the benefits of OOD evaluation.

Zhongchao Qian, Tyler L. Hayes, Kushal Kafle et al.

Traditionally, deep convolutional neural networks consist of a series of convolutional and pooling layers followed by one or more fully connected (FC) layers to perform the final classification. While this design has been successful, for datasets with a large number of categories, the fully connected layers often account for a large percentage of the network's parameters. For applications with memory constraints, such as mobile devices and embedded platforms, this is not ideal. Recently, a family of architectures that involve replacing the learned fully connected output layer with a fixed layer has been proposed as a way to achieve better efficiency. In this paper we examine this idea further and demonstrate that fixed classifiers offer no additional benefit compared to simply removing the output layer along with its parameters. We further demonstrate that the typical approach of having a fully connected final output layer is inefficient in terms of parameter count. We are able to achieve comparable performance to a traditionally learned fully connected classification output layer on the ImageNet-1K, CIFAR-100, Stanford Cars-196, and Oxford Flowers-102 datasets, while not having a fully connected output layer at all.