Matthew B. Blaschko

Enshu Liu, Junyi Zhu, Zinan Lin et al. · microsoft-research

The rapid advancement of large language models (LLMs) has led to architectures with billions to trillions of parameters, posing significant deployment challenges due to their substantial demands on memory, processing power, and energy consumption. Sparse Mixture-of-Experts (SMoE) architectures have emerged as a solution, activating only a subset of parameters per token, thereby achieving faster inference while maintaining performance. However, SMoE models still face limitations in broader deployment due to their large parameter counts and significant GPU memory requirements. In this work, we introduce a gradient-free evolutionary strategy named EEP (Efficient Expert P}runing) to enhance the pruning of experts in SMoE models. EEP relies solely on model inference (i.e., no gradient computation) and achieves greater sparsity while maintaining or even improving performance on downstream tasks. EEP can be used to reduce both the total number of experts (thus saving GPU memory) and the number of active experts (thus accelerating inference). For example, we demonstrate that pruning up to 75% of experts in Mixtral $8\times7$B-Instruct results in a substantial reduction in parameters with minimal performance loss. Remarkably, we observe improved performance on certain tasks, such as a significant increase in accuracy on the SQuAD dataset (from 53.4% to 75.4%), when pruning half of the experts. With these results, EEP not only lowers the barrier to deploying SMoE models,but also challenges the conventional understanding of model pruning by showing that fewer experts can lead to better task-specific performance without any fine-tuning. Code is available at https://github.com/imagination-research/EEP.

Zifu Wang, Maxim Berman, Amal Rannen-Triki et al.

Semantic segmentation datasets often exhibit two types of imbalance: \textit{class imbalance}, where some classes appear more frequently than others and \textit{size imbalance}, where some objects occupy more pixels than others. This causes traditional evaluation metrics to be biased towards \textit{majority classes} (e.g. overall pixel-wise accuracy) and \textit{large objects} (e.g. mean pixel-wise accuracy and per-dataset mean intersection over union). To address these shortcomings, we propose the use of fine-grained mIoUs along with corresponding worst-case metrics, thereby offering a more holistic evaluation of segmentation techniques. These fine-grained metrics offer less bias towards large objects, richer statistical information, and valuable insights into model and dataset auditing. Furthermore, we undertake an extensive benchmark study, where we train and evaluate 15 modern neural networks with the proposed metrics on 12 diverse natural and aerial segmentation datasets. Our benchmark study highlights the necessity of not basing evaluations on a single metric and confirms that fine-grained mIoUs reduce the bias towards large objects. Moreover, we identify the crucial role played by architecture designs and loss functions, which lead to best practices in optimizing fine-grained metrics. The code is available at \href{https://github.com/zifuwanggg/JDTLosses}{https://github.com/zifuwanggg/JDTLosses}.

Huy Hoang Nguyen, Matthew B. Blaschko, Simo Saarakkala et al.

Deep neural networks are often applied to medical images to automate the problem of medical diagnosis. However, a more clinically relevant question that practitioners usually face is how to predict the future trajectory of a disease. Current methods for prognosis or disease trajectory forecasting often require domain knowledge and are complicated to apply. In this paper, we formulate the prognosis prediction problem as a one-to-many prediction problem. Inspired by a clinical decision-making process with two agents -- a radiologist and a general practitioner -- we predict prognosis with two transformer-based components that share information with each other. The first transformer in this framework aims to analyze the imaging data, and the second one leverages its internal states as inputs, also fusing them with auxiliary clinical data. The temporal nature of the problem is modeled within the transformer states, allowing us to treat the forecasting problem as a multi-task classification, for which we propose a novel loss. We show the effectiveness of our approach in predicting the development of structural knee osteoarthritis changes and forecasting Alzheimer's disease clinical status directly from raw multi-modal data. The proposed method outperforms multiple state-of-the-art baselines with respect to performance and calibration, both of which are needed for real-world applications. An open-source implementation of our method is made publicly available at \url{https://github.com/Oulu-IMEDS/CLIMATv2}.

Zifu Wang, Xuefei Ning, Matthew B. Blaschko

Intersection over Union (IoU) losses are surrogates that directly optimize the Jaccard index. Leveraging IoU losses as part of the loss function have demonstrated superior performance in semantic segmentation tasks compared to optimizing pixel-wise losses such as the cross-entropy loss alone. However, we identify a lack of flexibility in these losses to support vital training techniques like label smoothing, knowledge distillation, and semi-supervised learning, mainly due to their inability to process soft labels. To address this, we introduce Jaccard Metric Losses (JMLs), which are identical to the soft Jaccard loss in standard settings with hard labels but are fully compatible with soft labels. We apply JMLs to three prominent use cases of soft labels: label smoothing, knowledge distillation and semi-supervised learning, and demonstrate their potential to enhance model accuracy and calibration. Our experiments show consistent improvements over the cross-entropy loss across 4 semantic segmentation datasets (Cityscapes, PASCAL VOC, ADE20K, DeepGlobe Land) and 13 architectures, including classic CNNs and recent vision transformers. Remarkably, our straightforward approach significantly outperforms state-of-the-art knowledge distillation and semi-supervised learning methods. The code is available at \href{https://github.com/zifuwanggg/JDTLosses}{https://github.com/zifuwanggg/JDTLosses}.

Zifu Wang, Teodora Popordanoska, Jeroen Bertels et al.

The soft Dice loss (SDL) has taken a pivotal role in numerous automated segmentation pipelines in the medical imaging community. Over the last years, some reasons behind its superior functioning have been uncovered and further optimizations have been explored. However, there is currently no implementation that supports its direct utilization in scenarios involving soft labels. Hence, a synergy between the use of SDL and research leveraging the use of soft labels, also in the context of model calibration, is still missing. In this work, we introduce Dice semimetric losses (DMLs), which (i) are by design identical to SDL in a standard setting with hard labels, but (ii) can be employed in settings with soft labels. Our experiments on the public QUBIQ, LiTS and KiTS benchmarks confirm the potential synergy of DMLs with soft labels (e.g. averaging, label smoothing, and knowledge distillation) over hard labels (e.g. majority voting and random selection). As a result, we obtain superior Dice scores and model calibration, which supports the wider adoption of DMLs in practice. The code is available at https://github.com/zifuwanggg/JDTLosses

Zifu Wang, Matthew B. Blaschko

UNet [27] is widely used in semantic segmentation due to its simplicity and effectiveness. However, its manually-designed architecture is applied to a large number of problem settings, either with no architecture optimizations, or with manual tuning, which is time consuming and can be sub-optimal. In this work, firstly, we propose Markov Random Field Neural Architecture Search (MRF-NAS) that extends and improves the recent Adaptive and Optimal Network Width Search (AOWS) method [4] with (i) a more general MRF framework (ii) diverse M-best loopy inference (iii) differentiable parameter learning. This provides the necessary NAS framework to efficiently explore network architectures that induce loopy inference graphs, including loops that arise from skip connections. With UNet as the backbone, we find an architecture, MRF-UNet, that shows several interesting characteristics. Secondly, through the lens of these characteristics, we identify the sub-optimality of the original UNet architecture and further improve our results with MRF-UNetV2. Experiments show that our MRF-UNets significantly outperform several benchmarks on three aerial image datasets and two medical image datasets while maintaining low computational costs. The code is available at: https://github.com/zifuwanggg/MRF-UNets.

Zehao Wang, Han Zhou, Matthew B. Blaschko et al.

Novel view synthesis of indoor scenes can be achieved by capturing a monocular video sequence of the environment. However, redundant information caused by artificial movements in the input video data reduces the efficiency of scene modeling. To address this, we formulate the problem as a combinatorial optimization task for view subset selection. In this work, we propose a novel subset selection framework that integrates a comprehensive diversity-based measurement with well-designed utility functions. We provide a theoretical analysis of these utility functions and validate their effectiveness through extensive experiments. Furthermore, we introduce IndoorTraj, a novel dataset designed for indoor novel view synthesis, featuring complex and extended trajectories that simulate intricate human behaviors. Experiments on IndoorTraj show that our framework consistently outperforms baseline strategies while using only 5-20% of the data, highlighting its remarkable efficiency and effectiveness. The code is available at: https://github.com/zehao-wang/IndoorTraj

Mang Ning, Mingxiao Li, Le Zhang et al.

In this paper, we study the diffusability (learnability) of variational autoencoders (VAE) in latent diffusion. First, we show that pixel-space diffusion trained with an MSE objective is inherently biased toward learning low and mid spatial frequencies, and that the power-law power spectral density (PSD) of natural images makes this bias perceptually beneficial. Motivated by this result, we propose the \emph{Spectrum Matching Hypothesis}: latents with superior diffusability should (i) follow a flattened power-law PSD (\emph{Encoding Spectrum Matching}, ESM) and (ii) preserve frequency-to-frequency semantic correspondence through the decoder (\emph{Decoding Spectrum Matching}, DSM). In practice, we apply ESM by matching the PSD between images and latents, and DSM via shared spectral masking with frequency-aligned reconstruction. Importantly, Spectrum Matching provides a unified view that clarifies prior observations of over-noisy or over-smoothed latents, and interprets several recent methods as special cases (e.g., VA-VAE, EQ-VAE). Experiments suggest that Spectrum Matching yields superior diffusion generation on CelebA and ImageNet datasets, and outperforms prior approaches. Finally, we extend the spectral view to representation alignment (REPA): we show that the directional spectral energy of the target representation is crucial for REPA, and propose a DoG-based method to further improve the performance of REPA. Our code is available https://github.com/forever208/SpectrumMatching.

Teodora Popordanoska, Raphael Sayer, Matthew B. Blaschko

Calibrated probabilistic classifiers are models whose predicted probabilities can directly be interpreted as uncertainty estimates. It has been shown recently that deep neural networks are poorly calibrated and tend to output overconfident predictions. As a remedy, we propose a low-bias, trainable calibration error estimator based on Dirichlet kernel density estimates, which asymptotically converges to the true $L_p$ calibration error. This novel estimator enables us to tackle the strongest notion of multiclass calibration, called canonical (or distribution) calibration, while other common calibration methods are tractable only for top-label and marginal calibration. The computational complexity of our estimator is $\mathcal{O}(n^2)$, the convergence rate is $\mathcal{O}(n^{-1/2})$, and it is unbiased up to $\mathcal{O}(n^{-2})$, achieved by a geometric series debiasing scheme. In practice, this means that the estimator can be applied to small subsets of data, enabling efficient estimation and mini-batch updates. The proposed method has a natural choice of kernel, and can be used to generate consistent estimates of other quantities based on conditional expectation, such as the sharpness of a probabilistic classifier. Empirical results validate the correctness of our estimator, and demonstrate its utility in canonical calibration error estimation and calibration error regularized risk minimization.

Jordy Van Landeghem, Sanket Biswas, Matthew B. Blaschko et al.

This paper highlights the need to bring document classification benchmarking closer to real-world applications, both in the nature of data tested ($X$: multi-channel, multi-paged, multi-industry; $Y$: class distributions and label set variety) and in classification tasks considered ($f$: multi-page document, page stream, and document bundle classification, ...). We identify the lack of public multi-page document classification datasets, formalize different classification tasks arising in application scenarios, and motivate the value of targeting efficient multi-page document representations. An experimental study on proposed multi-page document classification datasets demonstrates that current benchmarks have become irrelevant and need to be updated to evaluate complete documents, as they naturally occur in practice. This reality check also calls for more mature evaluation methodologies, covering calibration evaluation, inference complexity (time-memory), and a range of realistic distribution shifts (e.g., born-digital vs. scanning noise, shifting page order). Our study ends on a hopeful note by recommending concrete avenues for future improvements.}

Wangduo Xie, Matthew B. Blaschko

CT images corrupted by metal artifacts have serious negative effects on clinical diagnosis. Considering the difficulty of collecting paired data with ground truth in clinical settings, unsupervised methods for metal artifact reduction are of high interest. However, it is difficult for previous unsupervised methods to retain structural information from CT images while handling the non-local characteristics of metal artifacts. To address these challenges, we proposed a novel Dense Transformer based Enhanced Coding Network (DTEC-Net) for unsupervised metal artifact reduction. Specifically, we introduce a Hierarchical Disentangling Encoder, supported by the high-order dense process, and transformer to obtain densely encoded sequences with long-range correspondence. Then, we present a second-order disentanglement method to improve the dense sequence's decoding process. Extensive experiments and model discussions illustrate DTEC-Net's effectiveness, which outperforms the previous state-of-the-art methods on a benchmark dataset, and greatly reduces metal artifacts while restoring richer texture details.

Karel Moens, Matthew B. Blaschko, Tinne Tuytelaars et al.

Whole slide image (WSI) normalization remains a vital preprocessing step in computational pathology. Increasingly driven by deep learning, these models learn to approximate data distributions from training examples. This often results in outputs that gravitate toward the average, potentially masking diagnostically important features. More critically, they can introduce hallucinated content, artifacts that appear realistic but are not present in the original tissue, posing a serious threat to downstream analysis. These hallucinations are nearly impossible to detect visually, and current evaluation practices often overlook them. In this work, we demonstrate that the risk of hallucinations is real and underappreciated. While many methods perform adequately on public datasets, we observe a concerning frequency of hallucinations when these same models are retrained and evaluated on real-world clinical data. To address this, we propose a novel image comparison measure designed to automatically detect hallucinations in normalized outputs. Using this measure, we systematically evaluate several well-cited normalization methods retrained on real-world data, revealing significant inconsistencies and failures that are not captured by conventional metrics. Our findings underscore the need for more robust, interpretable normalization techniques and stricter validation protocols in clinical deployment.

Sinnu Susan Thomas, Jacopo Palandri, Mohsen Lakehal-ayat et al.

Engineering design is traditionally performed by hand: an expert makes design proposals based on past experience, and these proposals are then tested for compliance with certain target specifications. Testing for compliance is performed first by computer simulation using what is called a discipline model. Such a model can be implemented by a finite element analysis, multibody systems approach, etc. Designs passing this simulation are then considered for physical prototyping. The overall process may take months, and is a significant cost in practice. We have developed a Bayesian optimization system for partially automating this process by directly optimizing compliance with the target specification with respect to the design parameters. The proposed method is a general framework for computing a generalized inverse of a high-dimensional non-linear function that does not require e.g. gradient information, which is often unavailable from discipline models. We furthermore develop a two-tier convergence criterion based on (i) convergence to a solution optimally satisfying all specified design criteria, or (ii) convergence to a minimum-norm solution. We demonstrate the proposed approach on a vehicle chassis design problem motivated by an industry setting using a state-of-the-art commercial discipline model. We show that the proposed approach is general, scalable, and efficient, and that the novel convergence criteria can be implemented straightforwardly based on existing concepts and subroutines in popular Bayesian optimization software packages.

Han Zhou, Aida Ashrafi, Matthew B. Blaschko

Low-bit width neural networks have been extensively explored for deployment on edge devices to reduce computational resources. Existing approaches have focused on gradient-based optimization in a two-stage train-and-compress setting or as a combined optimization where gradients are quantized during training. Such schemes require high-performance hardware during the training phase and usually store an equivalent number of full-precision weights apart from the quantized weights. In this paper, we explore methods of direct combinatorial optimization in the problem of risk minimization with binary weights, which can be made equivalent to a non-monotone submodular maximization under certain conditions. We employ an approximation algorithm for the cases with single and multilayer neural networks. For linear models, it has $\mathcal{O}(nd)$ time complexity where $n$ is the sample size and $d$ is the data dimension. We show that a combination of greedy coordinate descent and this novel approach can attain competitive accuracy on binary classification tasks.

Teodora Popordanoska, Aleksei Tiulpin, Wacha Bounliphone et al.

The eigendecomposition of a matrix is the central procedure in probabilistic models based on matrix factorization, for instance principal component analysis and topic models. Quantifying the uncertainty of such a decomposition based on a finite sample estimate is essential to reasoning under uncertainty when employing such models. This paper tackles the challenge of computing confidence bounds on the individual entries of eigenvectors of a covariance matrix of fixed dimension. Moreover, we derive a method to bound the entries of the inverse covariance matrix, the so-called precision matrix. The assumptions behind our method are minimal and require that the covariance matrix exists, and its empirical estimator converges to the true covariance. We make use of the theory of U-statistics to bound the $L_2$ perturbation of the empirical covariance matrix. From this result, we obtain bounds on the eigenvectors using Weyl's theorem and the eigenvalue-eigenvector identity and we derive confidence intervals on the entries of the precision matrix using matrix inversion perturbation bounds. As an application of these results, we demonstrate a new statistical test, which allows us to test for non-zero values of the precision matrix. We compare this test to the well-known Fisher-z test for partial correlations, and demonstrate the soundness and scalability of the proposed statistical test, as well as its application to real-world data from medical and physics domains.

Pedro R. A. S. Bassi, Wenxuan Li, Yucheng Tang et al.

How can we test AI performance? This question seems trivial, but it isn't. Standard benchmarks often have problems such as in-distribution and small-size test sets, oversimplified metrics, unfair comparisons, and short-term outcome pressure. As a consequence, good performance on standard benchmarks does not guarantee success in real-world scenarios. To address these problems, we present Touchstone, a large-scale collaborative segmentation benchmark of 9 types of abdominal organs. This benchmark is based on 5,195 training CT scans from 76 hospitals around the world and 5,903 testing CT scans from 11 additional hospitals. This diverse test set enhances the statistical significance of benchmark results and rigorously evaluates AI algorithms across various out-of-distribution scenarios. We invited 14 inventors of 19 AI algorithms to train their algorithms, while our team, as a third party, independently evaluated these algorithms on three test sets. In addition, we also evaluated pre-existing AI frameworks--which, differing from algorithms, are more flexible and can support different algorithms--including MONAI from NVIDIA, nnU-Net from DKFZ, and numerous other open-source frameworks. We are committed to expanding this benchmark to encourage more innovation of AI algorithms for the medical domain.

Zifu Wang, Junyi Zhu, Bo Tang et al.

The application of rule-based reinforcement learning (RL) to multimodal large language models (MLLMs) introduces unique challenges and potential deviations from findings in text-only domains, particularly for perception-heavy tasks. This paper provides a comprehensive study of rule-based visual RL, using jigsaw puzzles as a structured experimental framework. Jigsaw puzzles offer inherent ground truth, adjustable difficulty, and demand complex decision-making, making them ideal for this study. Our research reveals several key findings: \textit{Firstly,} we find that MLLMs, initially performing near to random guessing on the simplest jigsaw puzzles, achieve near-perfect accuracy and generalize to complex, unseen configurations through fine-tuning. \textit{Secondly,} training on jigsaw puzzles can induce generalization to other visual tasks, with effectiveness tied to specific task configurations. \textit{Thirdly,} MLLMs can learn and generalize with or without explicit reasoning, though open-source models often favor direct answering. Consequently, even when trained for step-by-step reasoning, they can ignore the thinking process in deriving the final answer. \textit{Fourthly,} we observe that complex reasoning patterns appear to be pre-existing rather than emergent, with their frequency increasing alongside training and task difficulty. \textit{Finally,} our results demonstrate that RL exhibits more effective generalization than Supervised Fine-Tuning (SFT), and an initial SFT cold start phase can hinder subsequent RL optimization. Although these observations are based on jigsaw puzzles and may vary across other visual tasks, this research contributes a valuable piece of jigsaw to the larger puzzle of collective understanding rule-based visual RL and its potential in multimodal learning. The code is available at: https://github.com/zifuwanggg/Jigsaw-R1

Enshu Liu, Junyi Zhu, Zinan Lin et al. · microsoft-research

Diffusion Models (DM) and Consistency Models (CM) are two types of popular generative models with good generation quality on various tasks. When training DM and CM, intermediate weight checkpoints are not fully utilized and only the last converged checkpoint is used. In this work, we find that high-quality model weights often lie in a basin which cannot be reached by SGD but can be obtained by proper checkpoint averaging. Based on these observations, we propose LCSC, a simple but effective and efficient method to enhance the performance of DM and CM, by combining checkpoints along the training trajectory with coefficients deduced from evolutionary search. We demonstrate the value of LCSC through two use cases: $\textbf{(a) Reducing training cost.}$ With LCSC, we only need to train DM/CM with fewer number of iterations and/or lower batch sizes to obtain comparable sample quality with the fully trained model. For example, LCSC achieves considerable training speedups for CM (23$\times$ on CIFAR-10 and 15$\times$ on ImageNet-64). $\textbf{(b) Enhancing pre-trained models.}$ Assuming full training is already done, LCSC can further improve the generation quality or speed of the final converged models. For example, LCSC achieves better performance using 1 number of function evaluation (NFE) than the base model with 2 NFE on consistency distillation, and decreases the NFE of DM from 15 to 9 while maintaining the generation quality on CIFAR-10. Our code is available at https://github.com/imagination-research/LCSC.

Wangduo Xie, Richard Schoonhoven, Tristan van Leeuwen et al.

Computed tomography (CT) reconstruction plays a crucial role in industrial nondestructive testing and medical diagnosis. Sparse view CT reconstruction aims to reconstruct high-quality CT images while only using a small number of projections, which helps to improve the detection speed of industrial assembly lines and is also meaningful for reducing radiation in medical scenarios. Sparse CT reconstruction methods based on implicit neural representations (INRs) have recently shown promising performance, but still produce artifacts because of the difficulty of obtaining useful prior information. In this work, we incorporate a powerful prior: the total number of material categories of objects. To utilize the prior, we design AC-IND, a self-supervised method based on Attenuation Coefficient Estimation and Implicit Neural Distribution. Specifically, our method first transforms the traditional INR from scalar mapping to probability distribution mapping. Then we design a compact attenuation coefficient estimator initialized with values from a rough reconstruction and fast segmentation. Finally, our algorithm finishes the CT reconstruction by jointly optimizing the estimator and the generated distribution. Through experiments, we find that our method not only outperforms the comparative methods in sparse CT reconstruction but also can automatically generate semantic segmentation maps.

Jiameng Li, Minye Wu, Jiezhang Cao et al.

Long-form video understanding remains challenging for Video Large Language Models (VideoLLMs), as the dense frame sampling introduces massive visual tokens while sparse sampling risks missing critical temporal evidence and leading to LLM hallucination. Existing training-free token reduction methods either treat videos equally as static images or rely on segment-level merging heuristics, which weaken fine-grained spatiotemporal modeling and introduce additional overhead. In this paper, we propose EchoPrune, a lightweight and training-free token pruning method that improves temporal resolution under a fixed LLM-side visual token budget. Our core idea is to interpret redundant video tokens as temporal echoes: if a token is well reconstructed from the previous frame, it is merely a temporally redundant echo; otherwise, it may capture new events, motion, or query-relevant visual evidence. Based on this insight, EchoPrune scores visual tokens by (i) query-guided crossmodal relevance and (ii) temporal reconstruction error, measured by correspondence matching and echo matching across consecutive frames. The selected tokens preserve task-relevant cues and temporal novelty while suppressing predictable redundancy, allowing VideoLLMs to observe more frames without increasing the decoding budget. Extensive experiments on LLaVA-OV, Qwen2.5VL, and Qwen3VL across six video understanding benchmarks show that EchoPrune enables VideoLLMs to process up to 20x frames under the same token budget, yielding improved performance (+8.6%) and inference speedup (5.6x for prefilling) on Qwen2.5VL-7B.

Wangduo Xie, Matthew B. Blaschko

Computed Tomography (CT) plays a vital role in inspecting the internal structures of industrial objects. Furthermore, achieving high-quality CT reconstruction from sparse views is essential for reducing production costs. While classic implicit neural networks have shown promising results for sparse reconstruction, they are unable to leverage shape priors of objects. Motivated by the observation that numerous industrial objects exhibit rectangular structures, we propose a novel \textbf{N}eural \textbf{A}daptive \textbf{B}inning (\textbf{NAB}) method that effectively integrates rectangular priors into the reconstruction process. Specifically, our approach first maps coordinate space into a binned vector space. This mapping relies on an innovative binning mechanism based on differences between shifted hyperbolic tangent functions, with our extension enabling rotations around the input-plane normal vector. The resulting representations are then processed by a neural network to predict CT attenuation coefficients. This design enables end-to-end optimization of the encoding parameters -- including position, size, steepness, and rotation -- via gradient flow from the projection data, thus enhancing reconstruction accuracy. By adjusting the smoothness of the binning function, NAB can generalize to objects with more complex geometries. This research provides a new perspective on integrating shape priors into neural network-based reconstruction. Extensive experiments demonstrate that NAB achieves superior performance on two industrial datasets. It also maintains robust on medical datasets when the binning function is extended to more general expression. The code will be made available.

Gorjan Radevski, Teodora Popordanoska, Matthew B. Blaschko et al.

Audio-visual understanding is a rapidly evolving field that seeks to integrate and interpret information from both auditory and visual modalities. Despite recent advances in multi-modal learning, existing benchmarks often suffer from strong visual bias -- where answers can be inferred from visual data alone -- and provide only aggregate scores that conflate multiple sources of error. This makes it difficult to determine whether models struggle with visual understanding, audio interpretation, or audio-visual alignment. In this work, we introduce DAVE (Diagnostic Audio Visual Evaluation), a novel benchmark dataset designed to systematically evaluate audio-visual models across controlled challenges. DAVE alleviates existing limitations by (i) ensuring both modalities are necessary to answer correctly and (ii) decoupling evaluation into atomic subcategories. Our detailed analysis of state-of-the-art models reveals specific failure modes and provides targeted insights for improvement. By offering this standardized diagnostic framework, we aim to facilitate more robust development of audio-visual models. The dataset is released: https://github.com/gorjanradevski/dave

Junyi Zhu, Shuochen Liu, Yu Yu et al.

Large language models (LLMs) excel in generating coherent text, but they often struggle with context awareness, leading to inaccuracies in tasks requiring faithful adherence to provided information. We introduce FastMem, a novel method designed to enhance instruction fine-tuned LLMs' context awareness through fast memorization of the prompt. FastMem maximizes the likelihood of the prompt before inference by updating only the last Feed-Forward Network (FFN) module. This targeted approach ensures efficient optimization without overfitting, significantly improving the model's ability to comprehend and accurately follow the context. Our experiments demonstrate substantial gains in reading comprehension, text summarization and adherence to output structures. For instance, FastMem improves the accuracy of Llama 3-8B-Inst on the NQ-SWAP dataset from 59.1% to 71.6%, and reduces the output structure failure rate of Qwen 1.5-4B-Chat from 34.9% to 25.5%. Extensive experimental results highlight FastMem's potential to offer a robust solution to enhance the reliability and accuracy of LLMs in various applications. Our code is available at: https://github.com/IAAR-Shanghai/FastMem

Xuefei Ning, Zifu Wang, Shiyao Li et al.

Teaching to improve student models (e.g., knowledge distillation) is an extensively studied methodology in LLMs. However, for humans, teaching improves not only students but also teachers, by fostering more rigorous and clear reasoning as well as knowledge building. We ask: Can LLMs also learn by teaching (LbT) for better reasoning? If the answer is yes, we can potentially unlock the possibility of continuously advancing the models without solely relying on human-produced data or stronger models. In this paper, we provide a preliminary exploration on this question. We show that LbT ideas can be incorporated into existing LLM training/prompting pipelines and bring improvements. Specifically, we design three methods, each mimicking one of the three levels of LbT: observing students' feedback, learning from the feedback, and learning iteratively, with the goals of improving answer accuracy without training or improving models' inherent capability with fine-tuning. We reveal some findings: (1) Teaching materials that make it easier for students to learn have clearer and more accurate logic when using in-context learning as the student's "learning" method; (2) Weak-to-strong generalization: LbT might help improve strong models by teaching weak models; (3) Diversity in students might help: teaching multiple students could be better than teaching one student or the teacher itself. We hope that our exploration can inspire future research on LbT and more broadly adopting the advanced techniques in education to improve LLMs. The code and website are at https://github.com/imagination-research/lbt and https://sites.google.com/view/llm-learning-by-teaching.

Chang Tian, Matthew B. Blaschko, Wenpeng Yin et al.

Fine-grained category discovery using only coarse-grained supervision is a cost-effective yet challenging task. Previous training methods focus on aligning query samples with positive samples and distancing them from negatives. They often neglect intra-category and inter-category semantic similarities of fine-grained categories when navigating sample distributions in the embedding space. Furthermore, some evaluation techniques that rely on pre-collected test samples are inadequate for real-time applications. To address these shortcomings, we introduce a method that successfully detects fine-grained clusters of semantically similar texts guided by a novel objective function. The method uses semantic similarities in a logarithmic space to guide sample distributions in the Euclidean space and to form distinct clusters that represent fine-grained categories. We also propose a centroid inference mechanism to support real-time applications. The efficacy of the method is both theoretically justified and empirically confirmed on three benchmark tasks. The proposed objective function is integrated in multiple contrastive learning based neural models. Its results surpass existing state-of-the-art approaches in terms of Accuracy, Adjusted Rand Index and Normalized Mutual Information of the detected fine-grained categories. Code and data will be available at Code and data are publicly available at https://github.com/changtianluckyforever/F-grained-STAR.

Junyi Zhu, Ruicong Yao, Matthew B. Blaschko

In Federated Learning (FL) and many other distributed training frameworks, collaborators can hold their private data locally and only share the network weights trained with the local data after multiple iterations. Gradient inversion is a family of privacy attacks that recovers data from its generated gradients. Seemingly, FL can provide a degree of protection against gradient inversion attacks on weight updates, since the gradient of a single step is concealed by the accumulation of gradients over multiple local iterations. In this work, we propose a principled way to extend gradient inversion attacks to weight updates in FL, thereby better exposing weaknesses in the presumed privacy protection inherent in FL. In particular, we propose a surrogate model method based on the characteristic of two-dimensional gradient flow and low-rank property of local updates. Our method largely boosts the ability of gradient inversion attacks on weight updates containing many iterations and achieves state-of-the-art (SOTA) performance. Additionally, our method runs up to $100\times$ faster than the SOTA baseline in the common FL scenario. Our work re-evaluates and highlights the privacy risk of sharing network weights. Our code is available at https://github.com/JunyiZhu-AI/surrogate_model_extension.

Junyi Zhu, Xingchen Ma, Matthew B. Blaschko

Federated Learning (FL) is a distributed learning scheme to train a shared model across clients. One common and fundamental challenge in FL is that the sets of data across clients could be non-identically distributed and have different sizes. Personalized Federated Learning (PFL) attempts to solve this challenge via locally adapted models. In this work, we present a novel framework for PFL based on hierarchical Bayesian modeling and variational inference. A global model is introduced as a latent variable to augment the joint distribution of clients' parameters and capture the common trends of different clients, optimization is derived based on the principle of maximizing the marginal likelihood and conducted using variational expectation maximization. Our algorithm gives rise to a closed-form estimation of a confidence value which comprises the uncertainty of clients' parameters and local model deviations from the global model. The confidence value is used to weigh clients' parameters in the aggregation stage and adjust the regularization effect of the global model. We evaluate our method through extensive empirical studies on multiple datasets. Experimental results show that our approach obtains competitive results under mild heterogeneous circumstances while significantly outperforming state-of-the-art PFL frameworks in highly heterogeneous settings. Our code is available at https://github.com/JunyiZhu-AI/confidence_aware_PFL.

Aleksei Tiulpin, Matthew B. Blaschko

Ensembles of independently trained neural networks are a state-of-the-art approach to estimate predictive uncertainty in Deep Learning, and can be interpreted as an approximation of the posterior distribution via a mixture of delta functions. The training of ensembles relies on non-convexity of the loss landscape and random initialization of their individual members, making the resulting posterior approximation uncontrolled. This paper proposes a novel and principled method to tackle this limitation, minimizing an $f$-divergence between the true posterior and a kernel density estimator (KDE) in a function space. We analyze this objective from a combinatorial point of view, and show that it is submodular with respect to mixture components for any $f$. Subsequently, we consider the problem of greedy ensemble construction. From the marginal gain on the negative $f$-divergence, which quantifies an improvement in posterior approximation yielded by adding a new component into the KDE, we derive a novel diversity term for ensemble methods. The performance of our approach is demonstrated on computer vision out-of-distribution detection benchmarks in a range of architectures trained on multiple datasets. The source code of our method is made publicly available at https://github.com/Oulu-IMEDS/greedy_ensembles_training.

Huy Hoang Nguyen, Simo Saarakkala, Matthew B. Blaschko et al.

In medical applications, deep learning methods are built to automate diagnostic tasks. However, a clinically relevant question that practitioners usually face, is how to predict the future trajectory of a disease (prognosis). Current methods for such a problem often require domain knowledge, and are complicated to apply. In this paper, we formulate the prognosis prediction problem as a one-to-many forecasting problem from multimodal data. Inspired by a clinical decision-making process with two agents -- a radiologist and a general practitioner, we model a prognosis prediction problem with two transformer-based components that share information between each other. The first block in this model aims to analyze the imaging data, and the second block leverages the internal representations of the first one as inputs, also fusing them with auxiliary patient data. We show the effectiveness of our method in predicting the development of structural knee osteoarthritis changes over time. Our results show that the proposed method outperforms the state-of-the-art baselines in terms of various performance metrics. In addition, we empirically show that the existence of the multi-agent transformers with depths of 2 is sufficient to achieve good performances. Our code is publicly available at \url{https://github.com/MIPT-Oulu/CLIMAT}.

Mathijs Schuurmans, Maxim Berman, Matthew B. Blaschko

In this work, we evaluate the use of superpixel pooling layers in deep network architectures for semantic segmentation. Superpixel pooling is a flexible and efficient replacement for other pooling strategies that incorporates spatial prior information. We propose a simple and efficient GPU-implementation of the layer and explore several designs for the integration of the layer into existing network architectures. We provide experimental results on the IBSR and Cityscapes dataset, demonstrating that superpixel pooling can be leveraged to consistently increase network accuracy with minimal computational overhead. Source code is available at https://github.com/bermanmaxim/superpixPool

José Ignacio Orlando, Elena Prokofyeva, Mariana del Fresno et al.

Diabetic retinopathy is one of the leading causes of preventable blindness in the world. Its earliest sign are red lesions, a general term that groups both microaneurysms and hemorrhages. In daily clinical practice, these lesions are manually detected by physicians using fundus photographs. However, this task is tedious and time consuming, and requires an intensive effort due to the small size of the lesions and their lack of contrast. Computer-assisted diagnosis of DR based on red lesion detection is being actively explored due to its improvement effects both in clinicians consistency and accuracy. Several methods for detecting red lesions have been proposed in the literature, most of them based on characterizing lesion candidates using hand crafted features, and classifying them into true or false positive detections. Deep learning based approaches, by contrast, are scarce in this domain due to the high expense of annotating the lesions manually. In this paper we propose a novel method for red lesion detection based on combining both deep learned and domain knowledge. Features learned by a CNN are augmented by incorporating hand crafted features. Such ensemble vector of descriptors is used afterwards to identify true lesion candidates using a Random Forest classifier. We empirically observed that combining both sources of information significantly improve results with respect to using each approach separately. Furthermore, our method reported the highest performance on a per-lesion basis on DIARETDB1 and e-ophtha, and for screening and need for referral on MESSIDOR compared to a second human expert. Results highlight the fact that integrating manually engineered approaches with deep learned features is relevant to improve results when the networks are trained from lesion-level annotated data. An open source implementation of our system is publicly available online.

Amal Rannen Triki, Matthew B. Blaschko

Deep neural networks have had an enormous impact on image analysis. State-of-the-art training methods, based on weight decay and DropOut, result in impressive performance when a very large training set is available. However, they tend to have large problems overfitting to small data sets. Indeed, the available regularization methods deal with the complexity of the network function only indirectly. In this paper, we study the feasibility of directly using the $L_2$ function norm for regularization. Two methods to integrate this new regularization in the stochastic backpropagation are proposed. Moreover, the convergence of these new algorithms is studied. We finally show that they outperform the state-of-the-art methods in the low sample regime on benchmark datasets (MNIST and CIFAR10). The obtained results demonstrate very clear improvement, especially in the context of small sample regimes with data laying in a low dimensional manifold. Source code of the method can be found at \url{https://github.com/AmalRT/DNN_Reg}.

Teodora Popordanoska, Aleksei Tiulpin, Matthew B. Blaschko

Despite their impressive predictive performance in various computer vision tasks, deep neural networks (DNNs) tend to make overly confident predictions, which hinders their widespread use in safety-critical applications. While there have been recent attempts to calibrate DNNs, most of these efforts have primarily been focused on classification tasks, thus neglecting DNN-based object detectors. Although several recent works addressed calibration for object detection and proposed differentiable penalties, none of them are consistent estimators of established concepts in calibration. In this work, we tackle the challenge of defining and estimating calibration error specifically for this task. In particular, we adapt the definition of classification calibration error to handle the nuances associated with object detection, and predictions in structured output spaces more generally. Furthermore, we propose a consistent and differentiable estimator of the detection calibration error, utilizing kernel density estimation. Our experiments demonstrate the effectiveness of our estimator against competing train-time and post-hoc calibration methods, while maintaining similar detection performance.

Teodora Popordanoska, Sebastian G. Gruber, Aleksei Tiulpin et al.

Proper scoring rules evaluate the quality of probabilistic predictions, playing an essential role in the pursuit of accurate and well-calibrated models. Every proper score decomposes into two fundamental components -- proper calibration error and refinement -- utilizing a Bregman divergence. While uncertainty calibration has gained significant attention, current literature lacks a general estimator for these quantities with known statistical properties. To address this gap, we propose a method that allows consistent, and asymptotically unbiased estimation of all proper calibration errors and refinement terms. In particular, we introduce Kullback--Leibler calibration error, induced by the commonly used cross-entropy loss. As part of our results, we prove the relation between refinement and f-divergences, which implies information monotonicity in neural networks, regardless of which proper scoring rule is optimized. Our experiments validate empirically the claimed properties of the proposed estimator and suggest that the selection of a post-hoc calibration method should be determined by the particular calibration error of interest.

Jiayang Shi, Junyi Zhu, Daniel M. Pelt et al.

Computed Tomography (CT) is pivotal in industrial quality control and medical diagnostics. Sparse-view CT, offering reduced ionizing radiation, faces challenges due to its under-sampled nature, leading to ill-posed reconstruction problems. Recent advancements in Implicit Neural Representations (INRs) have shown promise in addressing sparse-view CT reconstruction. Recognizing that CT often involves scanning similar subjects, we propose a novel approach to improve reconstruction quality through joint reconstruction of multiple objects using INRs. This approach can potentially utilize the advantages of INRs and the common patterns observed across different objects. While current INR joint reconstruction techniques primarily focus on speeding up the learning process, they are not specifically tailored to enhance the final reconstruction quality. To address this gap, we introduce a novel INR-based Bayesian framework integrating latent variables to capture the common patterns across multiple objects under joint reconstruction. The common patterns then assist in the reconstruction of each object via latent variables, thereby improving the individual reconstruction. Extensive experiments demonstrate that our method achieves higher reconstruction quality with sparse views and remains robust to noise in the measurements as indicated by common numerical metrics. The obtained latent variables can also serve as network initialization for the new object and speed up the learning process.

Chang Tian, Matthew B. Blaschko, Mingzhe Xing et al.

Reinforcement learning (RL) has become a key technique for enhancing the reasoning abilities of large language models (LLMs), with policy-gradient algorithms dominating the post-training stage because of their efficiency and effectiveness. However, most existing benchmarks evaluate large-language-model reasoning under idealized settings, overlooking performance in realistic, non-ideal scenarios. We identify three representative non-ideal scenarios with practical relevance: summary inference, fine-grained noise suppression, and contextual filtering. We introduce a new research direction guided by brain-science findings that human reasoning remains reliable under imperfect inputs. We formally define and evaluate these challenging scenarios. We fine-tune three LLMs and a state-of-the-art large vision-language model (LVLM) using RL with a representative policy-gradient algorithm and then test their performance on eight public datasets. Our results reveal that while RL fine-tuning improves baseline reasoning under idealized settings, performance declines significantly across all three non-ideal scenarios, exposing critical limitations in advanced reasoning capabilities. Although we propose a scenario-specific remediation method, our results suggest current methods leave these reasoning deficits largely unresolved. This work highlights that the reasoning abilities of large models are often overstated and underscores the importance of evaluating models under non-ideal scenarios. The code and data will be released at XXXX.

Chang Tian, Mingzhe Xing, Zenglin Shi et al.

Predicting web service traffic has significant social value, as it can be applied to various practical scenarios, including but not limited to dynamic resource scaling, load balancing, system anomaly detection, service-level agreement compliance, and fraud detection. Web service traffic is characterized by frequent and drastic fluctuations over time and are influenced by heterogeneous web user behaviors, making accurate prediction a challenging task. Previous research has extensively explored statistical approaches, and neural networks to mine features from preceding service traffic time series for prediction. However, these methods have largely overlooked the causal relationships between services. Drawing inspiration from causality in ecological systems, we empirically recognize the causal relationships between web services. To leverage these relationships for improved web service traffic prediction, we propose an effective neural network module, CCMPlus, designed to extract causal relationship features across services. This module can be seamlessly integrated with existing time series models to consistently enhance the performance of web service traffic predictions. We theoretically justify that the causal correlation matrix generated by the CCMPlus module captures causal relationships among services. Empirical results on real-world datasets from Microsoft Azure, Alibaba Group, and Ant Group confirm that our method surpasses state-of-the-art approaches in Mean Squared Error (MSE) and Mean Absolute Error (MAE) for predicting service traffic time series. These findings highlight the efficacy of leveraging causal relationships for improved predictions.

Han Zhou, Jordy Van Landeghem, Teodora Popordanoska et al.

The selective classifier (SC) has been proposed for rank based uncertainty thresholding, which could have applications in safety critical areas such as medical diagnostics, autonomous driving, and the justice system. The Area Under the Risk-Coverage Curve (AURC) has emerged as the foremost evaluation metric for assessing the performance of SC systems. In this work, we present a formal statistical formulation of population AURC, presenting an equivalent expression that can be interpreted as a reweighted risk function. Through Monte Carlo methods, we derive empirical AURC plug-in estimators for finite sample scenarios. The weight estimators associated with these plug-in estimators are shown to be consistent, with low bias and tightly bounded mean squared error (MSE). The plug-in estimators are proven to converge at a rate of $\mathcal{O}(\sqrt{\ln(n)/n})$ demonstrating statistical consistency. We empirically validate the effectiveness of our estimators through experiments across multiple datasets, model architectures, and confidence score functions (CSFs), demonstrating consistency and effectiveness in fine-tuning AURC performance.

Abhishek Jha, Matthew B. Blaschko, Yuki M. Asano et al.

Last couple of years have witnessed a tremendous progress in self-supervised learning (SSL), the success of which can be attributed to the introduction of useful inductive biases in the learning process to learn meaningful visual representations while avoiding collapse. These inductive biases and constraints manifest themselves in the form of different optimization formulations in the SSL techniques, e.g. by utilizing negative examples in a contrastive formulation, or exponential moving average and predictor in BYOL and SimSiam. In this paper, we provide a framework to explain the stability mechanism of these different SSL techniques: i) we discuss the working mechanism of contrastive techniques like SimCLR, non-contrastive techniques like BYOL, SWAV, SimSiam, Barlow Twins, and DINO; ii) we provide an argument that despite different formulations these methods implicitly optimize a similar objective function, i.e. minimizing the magnitude of the expected representation over all data samples, or the mean of the data distribution, while maximizing the magnitude of the expected representation of individual samples over different data augmentations; iii) we provide mathematical and empirical evidence to support our framework. We formulate different hypotheses and test them using the Imagenet100 dataset.

Jiameng Li, Aleksei Tiulpin, Matthew B. Blaschko

For multimodal large language models (MLLMs), visual information is relatively sparse compared with text. As a result, research on visual pruning emerges for efficient inference. Current approaches typically measure token importance based on the attention scores in the visual encoder or in the LLM decoder, then select visual tokens with high attention scores while pruning others. In this paper, we pursue a different and more surgical approach. Instead of relying on mechanism-specific signals, we directly compute Mutual Information (MI) between visual and textual features themselves, prior to their interaction. This allows us to explicitly measure crossmodal dependency at the feature levels. Our MI-Pruner is simple, efficient and non-intrusive, requiring no access to internal attention maps or architectural modifications. Experimental results demonstrate that our approach outperforms previous attention-based pruning methods with minimal latency.

Teodora Popordanoska, Gorjan Radevski, Tinne Tuytelaars et al.

In the face of dataset shift, model calibration plays a pivotal role in ensuring the reliability of machine learning systems. Calibration error (CE) is an indicator of the alignment between the predicted probabilities and the classifier accuracy. While prior works have delved into the implications of dataset shift on calibration, existing CE estimators assume access to labels from the target domain, which are often unavailable in practice, i.e., when the model is deployed and used. This work addresses such challenging scenario, and proposes a novel CE estimator under label shift, which is characterized by changes in the marginal label distribution $p(Y)$, while keeping the conditional $p(X|Y)$ constant between the source and target distributions. Our contribution is an approach, which, by leveraging importance re-weighting of the labeled source distribution, provides consistent and asymptotically unbiased CE estimation with respect to the shifted target distribution. Empirical results across diverse real-world datasets, under various conditions and label-shift intensities, demonstrate the effectiveness and reliability of the proposed estimator.

Teodora Popordanoska, Jiameng Li, Matthew B. Blaschko

Contradictory multimodal inputs are common in real-world settings, yet existing benchmarks typically assume input consistency and fail to evaluate cross-modal contradiction detection - a fundamental capability for preventing hallucinations and ensuring reliability. We introduce CLASH, a novel benchmark for multimodal contradiction detection, featuring COCO images paired with contradictory captions containing controlled object-level or attribute-level contradictions. The samples include targeted questions evaluated in both multiple-choice and open-ended formats. The benchmark provides an extensive fine-tuning set filtered through automated quality checks, alongside a smaller human-verified diagnostic set. Our analysis of state-of-the-art models reveals substantial limitations in recognizing cross-modal conflicts, exposing systematic modality biases and category-specific weaknesses. Furthermore, we empirically demonstrate that targeted fine-tuning on CLASH substantially enhances conflict detection capabilities.

Dongli Xu, Aleksei Tiulpin, Matthew B. Blaschko

Autoregressive (AR) models have emerged as powerful tools for image generation by modeling images as sequences of discrete tokens. While Classifier-Free Guidance (CFG) has been adopted to improve conditional generation, its application in AR models faces two key issues: guidance diminishing, where the conditional-unconditional gap quickly vanishes as decoding progresses, and over-guidance, where strong conditions distort visual coherence. To address these challenges, we propose SoftCFG, an uncertainty-guided inference method that distributes adaptive perturbations across all tokens in the sequence. The key idea behind SoftCFG is to let each generated token contribute certainty-weighted guidance, ensuring that the signal persists across steps while resolving conflicts between text guidance and visual context. To further stabilize long-sequence generation, we introduce Step Normalization, which bounds cumulative perturbations of SoftCFG. Our method is training-free, model-agnostic, and seamlessly integrates with existing AR pipelines. Experiments show that SoftCFG significantly improves image quality over standard CFG and achieves state-of-the-art FID on ImageNet 256*256 among autoregressive models.

Huy Hoang Nguyen, Han Zhou, Matthew B. Blaschko et al.

Bayesian optimization with Gaussian processes (GP) is commonly used to optimize black-box functions. The Matérn and the Radial Basis Function (RBF) covariance functions are used frequently, but they do not make any assumptions about the domain of the function, which may limit their applicability in bounded domains. To address the limitation, we introduce the Beta kernel, a non-stationary kernel induced by a product of Beta distribution density functions. Such a formulation allows our kernel to naturally model functions on bounded domains. We present statistical evidence supporting the hypothesis that the kernel exhibits an exponential eigendecay rate, based on empirical analyses of its spectral properties across different settings. Our experimental results demonstrate the robustness of the Beta kernel in modeling functions with optima located near the faces or vertices of the unit hypercube. The experiments show that our kernel consistently outperforms a wide range of kernels, including the well-known Matérn and RBF, in different problems, including synthetic function optimization and the compression of vision and language models.

Han Zhou, Sebastian G. Gruber, Teodora Popordanoska et al.

Several variants of reweighted risk functionals, such as focal loss, inverse focal loss, and the Area Under the Risk--Coverage Curve (AURC), have been proposed for improving model calibration, yet their theoretical connections to calibration errors remain unclear. In this paper, we revisit a broad class of weighted risk functions commonly used in deep learning and establish a principled connection between calibration error and selective classification. We show that minimizing calibration error is closely linked to the selective classification paradigm and demonstrate that optimizing selective risk in low-confidence region naturally leads to improved calibration. This loss shares a similar reweighting strategy with dual focal loss but offers greater flexibility through the choice of confidence score functions (CSFs). Our approach uses a bin-based cumulative distribution function (CDF) approximation, enabling efficient gradient-based optimization without requiring expensive sorting and achieving $O(nK)$ complexity. Empirical evaluations demonstrate that our method achieves competitive calibration performance across a range of datasets and model architectures.

Jiameng Li, Teodora Popordanoska, Aleksei Tiulpin et al.

Ratio-based biomarkers -- such as the proportion of necrotic tissue within a tumor -- are widely used in clinical practice to support diagnosis, prognosis, and treatment planning. These biomarkers are typically estimated from soft segmentation outputs by computing region-wise ratios. Despite the high-stakes nature of clinical decision making, existing methods provide only point estimates, offering no measure of uncertainty. In this work, we propose a unified confidence-aware framework for estimating ratio-based biomarkers. Our uncertainty analysis stems from two observations: i) the probability ratio estimator inherently admits a statistical confidence interval regarding local randomness (bias and variance), ii) the segmentation network is not perfectly calibrated. We conduct a systematic analysis of error propagation in the segmentation-to-biomarker pipeline and identify model miscalibration as the dominant source of uncertainty. We leverage tunable parameters to control the confidence level of the derived bounds, allowing adaptation towards clinical practice. Extensive experiments show that our method produces statistically sound confidence intervals, with tunable confidence levels, enabling more trustworthy application of predictive biomarkers in clinical workflows.

Junyi Zhu, Ruicong Yao, Taha Ceritli et al.

Current network training paradigms primarily focus on either centralized or decentralized data regimes. However, in practice, data availability often exhibits a hybrid nature, where both regimes coexist. This hybrid setting presents new opportunities for model training, as the two regimes offer complementary trade-offs: decentralized data is abundant but subject to heterogeneity and communication constraints, while centralized data, though limited in volume and potentially unrepresentative, enables better curation and high-throughput access. Despite its potential, effectively combining these paradigms remains challenging, and few frameworks are tailored to hybrid data regimes. To address this, we propose a novel framework that constructs a model atlas from decentralized models and leverages centralized data to refine a global model within this structured space. The refined model is then used to reinitialize the decentralized models. Our method synergizes federated learning (to exploit decentralized data) and model merging (to utilize centralized data), enabling effective training under hybrid data availability. Theoretically, we show that our approach achieves faster convergence than methods relying solely on decentralized data, due to variance reduction in the merging process. Extensive experiments demonstrate that our framework consistently outperforms purely centralized, purely decentralized, and existing hybrid-adaptable methods. Notably, our method remains robust even when the centralized and decentralized data domains differ or when decentralized data contains noise, significantly broadening its applicability.

Mingshi Li, Dusan Grujicic, Ben Somers et al.

Remote sensing imagery from systems such as Sentinel provides full coverage of the Earth's surface at around 10-meter resolution. The remote sensing community has transitioned to extensive use of deep learning models due to their high performance on benchmarks such as the UCMerced and ISPRS Vaihingen datasets. Convolutional models such as UNet and ResNet variations are commonly employed for remote sensing but typically only accept three channels, as they were developed for RGB imagery, while satellite systems provide more than ten. Recently, several transformer architectures have been proposed for remote sensing, but they have not been extensively benchmarked and are typically used on small datasets such as Salinas Valley. Meanwhile, it is becoming feasible to obtain dense spatial land-use labels for entire first-level administrative divisions of some countries. Scaling law observations suggest that substantially larger multi-spectral transformer models could provide a significant leap in remote sensing performance in these settings. In this work, we propose ChromaFormer, a family of multi-spectral transformer models, which we evaluate across orders of magnitude differences in model parameters to assess their performance and scaling effectiveness on a densely labeled imagery dataset of Flanders, Belgium, covering more than 13,500 km^2 and containing 15 classes. We propose a novel multi-spectral attention strategy and demonstrate its effectiveness through ablations. Furthermore, we show that models many orders of magnitude larger than conventional architectures, such as UNet, lead to substantial accuracy improvements: a UNet++ model with 23M parameters achieves less than 65% accuracy, while a multi-spectral transformer with 655M parameters achieves over 95% accuracy on the Biological Valuation Map of Flanders.

Mingshi Li, Dusan Grujicic, Steven De Saeger et al.

In recent years, machine learning has become crucial in remote sensing analysis, particularly in the domain of Land-use/Land-cover (LULC). The synergy of machine learning and satellite imagery analysis has demonstrated significant productivity in this field, as evidenced by several studies. A notable challenge within this area is the semantic segmentation mapping of land usage over extensive territories, where the accessibility of accurate land-use data and the reliability of ground truth land-use labels pose significant difficulties. For example, providing a detailed and accurate pixel-wise labeled dataset of the Flanders region, a first-level administrative division of Belgium, can be particularly insightful. Yet there is a notable lack of regulated, formalized datasets and workflows for such studies in many regions globally. This paper introduces a comprehensive approach to addressing these gaps. We present a densely labeled ground truth map of Flanders paired with Sentinel-2 satellite imagery. Our methodology includes a formalized dataset division and sampling method, utilizing the topographic map layout 'Kaartbladversnijdingen,' and a detailed semantic segmentation model training pipeline. Preliminary benchmarking results are also provided to demonstrate the efficacy of our approach.

Teodora Popordanoska, Jeroen Bertels, Dirk Vandermeulen et al.

Machine learning driven medical image segmentation has become standard in medical image analysis. However, deep learning models are prone to overconfident predictions. This has led to a renewed focus on calibrated predictions in the medical imaging and broader machine learning communities. Calibrated predictions are estimates of the probability of a label that correspond to the true expected value of the label conditioned on the confidence. Such calibrated predictions have utility in a range of medical imaging applications, including surgical planning under uncertainty and active learning systems. At the same time it is often an accurate volume measurement that is of real importance for many medical applications. This work investigates the relationship between model calibration and volume estimation. We demonstrate both mathematically and empirically that if the predictor is calibrated per image, we can obtain the correct volume by taking an expectation of the probability scores per pixel/voxel of the image. Furthermore, we show that convex combinations of calibrated classifiers preserve volume estimation, but do not preserve calibration. Therefore, we conclude that having a calibrated predictor is a sufficient, but not necessary condition for obtaining an unbiased estimate of the volume. We validate our theoretical findings empirically on a collection of 18 different (calibrated) training strategies on the tasks of glioma volume estimation on BraTS 2018, and ischemic stroke lesion volume estimation on ISLES 2018 datasets.