Julia E. Vogt

Yuge Shi, Imant Daunhawer, Julia E. Vogt et al. · oxford

The robustness of machine learning algorithms to distributions shift is primarily discussed in the context of supervised learning (SL). As such, there is a lack of insight on the robustness of the representations learned from unsupervised methods, such as self-supervised learning (SSL) and auto-encoder based algorithms (AE), to distribution shift. We posit that the input-driven objectives of unsupervised algorithms lead to representations that are more robust to distribution shift than the target-driven objective of SL. We verify this by extensively evaluating the performance of SSL and AE on both synthetic and realistic distribution shift datasets. Following observations that the linear layer used for classification itself can be susceptible to spurious correlations, we evaluate the representations using a linear head trained on a small amount of out-of-distribution (OOD) data, to isolate the robustness of the learned representations from that of the linear head. We also develop "controllable" versions of existing realistic domain generalisation datasets with adjustable degrees of distribution shifts. This allows us to study the robustness of different learning algorithms under versatile yet realistic distribution shift conditions. Our experiments show that representations learned from unsupervised learning algorithms generalise better than SL under a wide variety of extreme as well as realistic distribution shifts.

Paweł Czyż, Frederic Grabowski, Julia E. Vogt et al. · eth-zurich

Mutual information is a general statistical dependency measure which has found applications in representation learning, causality, domain generalization and computational biology. However, mutual information estimators are typically evaluated on simple families of probability distributions, namely multivariate normal distribution and selected distributions with one-dimensional random variables. In this paper, we show how to construct a diverse family of distributions with known ground-truth mutual information and propose a language-independent benchmarking platform for mutual information estimators. We discuss the general applicability and limitations of classical and neural estimators in settings involving high dimensions, sparse interactions, long-tailed distributions, and high mutual information. Finally, we provide guidelines for practitioners on how to select appropriate estimator adapted to the difficulty of problem considered and issues one needs to consider when applying an estimator to a new data set.

Alexander Immer, Christoph Schultheiss, Julia E. Vogt et al.

We study the class of location-scale or heteroscedastic noise models (LSNMs), in which the effect $Y$ can be written as a function of the cause $X$ and a noise source $N$ independent of $X$, which may be scaled by a positive function $g$ over the cause, i.e., $Y = f(X) + g(X)N$. Despite the generality of the model class, we show the causal direction is identifiable up to some pathological cases. To empirically validate these theoretical findings, we propose two estimators for LSNMs: an estimator based on (non-linear) feature maps, and one based on neural networks. Both model the conditional distribution of $Y$ given $X$ as a Gaussian parameterized by its natural parameters. When the feature maps are correctly specified, we prove that our estimator is jointly concave, and a consistent estimator for the cause-effect identification task. Although the the neural network does not inherit those guarantees, it can fit functions of arbitrary complexity, and reaches state-of-the-art performance across benchmarks.

Ričards Marcinkevičs, Patricia Reis Wolfertstetter, Ugne Klimiene et al. · eth-zurich

Appendicitis is among the most frequent reasons for pediatric abdominal surgeries. Previous decision support systems for appendicitis have focused on clinical, laboratory, scoring, and computed tomography data and have ignored abdominal ultrasound, despite its noninvasive nature and widespread availability. In this work, we present interpretable machine learning models for predicting the diagnosis, management and severity of suspected appendicitis using ultrasound images. Our approach utilizes concept bottleneck models (CBM) that facilitate interpretation and interaction with high-level concepts understandable to clinicians. Furthermore, we extend CBMs to prediction problems with multiple views and incomplete concept sets. Our models were trained on a dataset comprising 579 pediatric patients with 1709 ultrasound images accompanied by clinical and laboratory data. Results show that our proposed method enables clinicians to utilize a human-understandable and intervenable predictive model without compromising performance or requiring time-consuming image annotation when deployed. For predicting the diagnosis, the extended multiview CBM attained an AUROC of 0.80 and an AUPR of 0.92, performing comparably to similar black-box neural networks trained and tested on the same dataset.

Paweł Czyż, Frederic Grabowski, Julia E. Vogt et al. · eth-zurich

The pointwise mutual information profile, or simply profile, is the distribution of pointwise mutual information for a given pair of random variables. One of its important properties is that its expected value is precisely the mutual information between these random variables. In this paper, we analytically describe the profiles of multivariate normal distributions and introduce a novel family of distributions, Bend and Mix Models, for which the profile can be accurately estimated using Monte Carlo methods. We then show how Bend and Mix Models can be used to study the limitations of existing mutual information estimators, investigate the behavior of neural critics used in variational estimators, and understand the effect of experimental outliers on mutual information estimation. Finally, we show how Bend and Mix Models can be used to obtain model-based Bayesian estimates of mutual information, suitable for problems with available domain expertise in which uncertainty quantification is necessary.

Irene Cannistraci, Simone Antonelli, Emanuele Palumbo et al.

Foundation models achieve state-of-the-art performance across different tasks, but their size and computational demands raise concerns about accessibility and sustainability. Existing efficiency methods often require additional retraining or finetuning, limiting their practicality. Recent findings suggest that deep neural networks exhibit internal representation similarities. While such similarities across different models have been exploited for enabling techniques such as model stitching and merging, intra-network redundancy remains underexplored as a source for efficiency gains. In this paper, we introduce Transformer Optimization using Adaptive and Simple Transformations (TOAST), a framework that exploits these redundancies to approximate entire transformer blocks with lightweight closed-form mappings, such as linear transformations or even the identity function, without any additional training. Across state-of-the-art pretrained vision models (e.g., ViT, DINOv2, DeiT) and datasets ranging from MNIST to ImageNet-1k, TOAST reduces parameters and computation while preserving, and in some cases improving, downstream performance. These results show that large portions of transformer depth can be replaced by trivial functions, opening a new perspective on efficient foundation models.

Ričards Marcinkevičs, Ece Ozkan, Julia E. Vogt · eth-zurich

Deep neural networks for image-based screening and computer-aided diagnosis have achieved expert-level performance on various medical imaging modalities, including chest radiographs. Recently, several works have indicated that these state-of-the-art classifiers can be biased with respect to sensitive patient attributes, such as race or gender, leading to growing concerns about demographic disparities and discrimination resulting from algorithmic and model-based decision-making in healthcare. Fair machine learning has focused on mitigating such biases against disadvantaged or marginalised groups, mainly concentrating on tabular data or natural images. This work presents two novel intra-processing techniques based on fine-tuning and pruning an already-trained neural network. These methods are simple yet effective and can be readily applied post hoc in a setting where the protected attribute is unknown during the model development and test time. In addition, we compare several intra- and post-processing approaches applied to debiasing deep chest X-ray classifiers. To the best of our knowledge, this is one of the first efforts studying debiasing methods on chest radiographs. Our results suggest that the considered approaches successfully mitigate biases in fully connected and convolutional neural networks offering stable performance under various settings. The discussed methods can help achieve group fairness of deep medical image classifiers when deploying them in domains with different fairness considerations and constraints.

Imant Daunhawer, Alice Bizeul, Emanuele Palumbo et al.

Contrastive learning is a cornerstone underlying recent progress in multi-view and multimodal learning, e.g., in representation learning with image/caption pairs. While its effectiveness is not yet fully understood, a line of recent work reveals that contrastive learning can invert the data generating process and recover ground truth latent factors shared between views. In this work, we present new identifiability results for multimodal contrastive learning, showing that it is possible to recover shared factors in a more general setup than the multi-view setting studied previously. Specifically, we distinguish between the multi-view setting with one generative mechanism (e.g., multiple cameras of the same type) and the multimodal setting that is characterized by distinct mechanisms (e.g., cameras and microphones). Our work generalizes previous identifiability results by redefining the generative process in terms of distinct mechanisms with modality-specific latent variables. We prove that contrastive learning can block-identify latent factors shared between modalities, even when there are nontrivial dependencies between factors. We empirically verify our identifiability results with numerical simulations and corroborate our findings on a complex multimodal dataset of image/text pairs. Zooming out, our work provides a theoretical basis for multimodal representation learning and explains in which settings multimodal contrastive learning can be effective in practice.

Ričards Marcinkevičs, Ece Ozkan, Julia E. Vogt · eth-zurich

Many modern research fields increasingly rely on collecting and analysing massive, often unstructured, and unwieldy datasets. Consequently, there is growing interest in machine learning and artificial intelligence applications that can harness this `data deluge'. This broad nontechnical overview provides a gentle introduction to machine learning with a specific focus on medical and biological applications. We explain the common types of machine learning algorithms and typical tasks that can be solved, illustrating the basics with concrete examples from healthcare. Lastly, we provide an outlook on open challenges, limitations, and potential impacts of machine-learning-powered medicine.

Kacper Sokol, Julia E. Vogt

Ante-hoc interpretability has become the holy grail of explainable artificial intelligence for high-stakes domains such as healthcare; however, this notion is elusive, lacks a widely-accepted definition and depends on the operational context. It can refer to predictive models whose structure adheres to domain-specific constraints, or ones that are inherently transparent. The latter conceptualisation assumes observers who judge this quality, whereas the former presupposes them to have technical and domain expertise (thus alienating other groups of explainees). Additionally, the distinction between ante-hoc interpretability and the less desirable post-hoc explainability, which refers to methods that construct a separate explanatory model, is vague given that transparent predictive models may still require (post-)processing to yield suitable explanatory insights. Ante-hoc interpretability is thus an overloaded concept that comprises a range of implicit properties, which we unpack in this paper to better understand what is needed for its safe adoption across high-stakes domains. To this end, we outline modelling and explaining desiderata that allow us to navigate its distinct realisations in view of the envisaged application and audience.

Alain Ryser, Laura Manduchi, Fabian Laumer et al.

We propose a novel anomaly detection method for echocardiogram videos. The introduced method takes advantage of the periodic nature of the heart cycle to learn three variants of a variational latent trajectory model (TVAE). While the first two variants (TVAE-C and TVAE-R) model strict periodic movements of the heart, the third (TVAE-S) is more general and allows shifts in the spatial representation throughout the video. All models are trained on the healthy samples of a novel in-house dataset of infant echocardiogram videos consisting of multiple chamber views to learn a normative prior of the healthy population. During inference, maximum a posteriori (MAP) based anomaly detection is performed to detect out-of-distribution samples in our dataset. The proposed method reliably identifies severe congenital heart defects, such as Ebstein's Anomaly or Shone-complex. Moreover, it achieves superior performance over MAP-based anomaly detection with standard variational autoencoders when detecting pulmonary hypertension and right ventricular dilation. Finally, we demonstrate that the proposed method enables interpretable explanations of its output through heatmaps highlighting the regions corresponding to anomalous heart structures.

Thomas M. Sutter, Laura Manduchi, Alain Ryser et al.

Partitioning a set of elements into subsets of a priori unknown sizes is essential in many applications. These subset sizes are rarely explicitly learned - be it the cluster sizes in clustering applications or the number of shared versus independent generative latent factors in weakly-supervised learning. Probability distributions over correct combinations of subset sizes are non-differentiable due to hard constraints, which prohibit gradient-based optimization. In this work, we propose the differentiable hypergeometric distribution. The hypergeometric distribution models the probability of different group sizes based on their relative importance. We introduce reparameterizable gradients to learn the importance between groups and highlight the advantage of explicitly learning the size of subsets in two typical applications: weakly-supervised learning and clustering. In both applications, we outperform previous approaches, which rely on suboptimal heuristics to model the unknown size of groups.

Ece Ozkan, Thomas M. Sutter, Yurong Hu et al.

Early detection of cardiac dysfunction through routine screening is vital for diagnosing cardiovascular diseases. An important metric of cardiac function is the left ventricular ejection fraction (EF), where lower EF is associated with cardiomyopathy. Echocardiography is a popular diagnostic tool in cardiology, with ultrasound being a low-cost, real-time, and non-ionizing technology. However, human assessment of echocardiograms for calculating EF is time-consuming and expertise-demanding, raising the need for an automated approach. In this work, we propose using the M(otion)-mode of echocardiograms for estimating the EF and classifying cardiomyopathy. We generate multiple artificial M-mode images from a single echocardiogram and combine them using off-the-shelf model architectures. Additionally, we extend contrastive learning (CL) to cardiac imaging to learn meaningful representations from exploiting structures in unlabeled data allowing the model to achieve high accuracy, even with limited annotations. Our experiments show that the supervised setting converges with only ten modes and is comparable to the baseline method while bypassing its cumbersome training process and being computationally much more efficient. Furthermore, CL using M-mode images is helpful for limited data scenarios, such as having labels for only 200 patients, which is common in medical applications.

Simon Böhi, Irene Cannistraci, Sergio Muñoz Gonzalez et al.

Echocardiography is a widely used modality for cardiac assessment due to its non-invasive and cost-effective nature, but the sparse and heterogeneous spatiotemporal views of the heart pose distinct challenges. Existing masked autoencoder (MAE) approaches typically process images or short clips independently, failing to capture the inherent multi-view structure required for coherent cardiac representation. We introduce Latent Attention Masked Autoencoder (LAMAE), a foundation model architecture tailored to the multi-view nature of medical imaging. LAMAE augments the standard MAE with a latent attention module that enables information exchange across frames and views directly in latent space. This allows the model to aggregate variable-length sequences and distinct views, reconstructing a holistic representation of cardiac function from partial observations. We pretrain LAMAE on MIMIC-IV-ECHO, a large-scale, uncurated dataset reflecting real-world clinical variability. To the best of our knowledge, we present the first results for predicting ICD-10 codes from MIMIC-IV-ECHO videos. Furthermore, we empirically demonstrate that representations learned from adult data transfer effectively to pediatric cohorts despite substantial anatomical differences. These results provide evidence that incorporating structural priors, such as multi-view attention, yields significantly more robust and transferable representations.

Jorge da Silva Goncalves, Laura Manduchi, Moritz Vandenhirtz et al.

This paper introduces Diffuse-TreeVAE, a deep generative model that integrates hierarchical clustering into the framework of Denoising Diffusion Probabilistic Models (DDPMs). The proposed approach generates new images by sampling from a root embedding of a learned latent tree VAE-based structure, it then propagates through hierarchical paths, and utilizes a second-stage DDPM to refine and generate distinct, high-quality images for each data cluster. The result is a model that not only improves image clarity but also ensures that the generated samples are representative of their respective clusters, addressing the limitations of previous VAE-based methods and advancing the state of clustering-based generative modeling.

Moritz Vandenhirtz, Samuel Ruipérez-Campillo, Simon Böhi et al.

Electrocardiograms (ECGs) are among the most widely available clinical signals and play a central role in cardiovascular diagnosis. While recent foundation models (FMs) have shown promise for learning transferable ECG representations, most existing pretraining approaches treat leads as independent channels and fail to explicitly leverage their strong structural redundancy. We introduce the latent attention masked autoencoder (LAMAE) FM that directly exploits this structure by learning cross-lead connection mechanisms during self-supervised pretraining. Our approach models higher-order interactions across leads through latent attention, enabling permutation-invariant aggregation and adaptive weighting of lead-specific representations. We provide empirical evidence on the Mimic-IV-ECG database that leveraging the cross-lead connection constitutes an effective form of structural supervision, improving representation quality and transferability. Our method shows strong performance in predicting ICD-10 codes, outperforming independent-lead masked modeling and alignment-based baselines.

Benjamin Gundersen, Nicolas Deperrois, Samuel Ruiperez-Campillo et al.

Recent advances in vision-language models (VLMs) have improved Chest X-ray (CXR) interpretation in multiple aspects. However, many medical VLMs rely solely on supervised fine-tuning (SFT), which optimizes next-token prediction without evaluating answer quality. In contrast, reinforcement learning (RL) can incorporate task-specific feedback, and its combination with explicit intermediate reasoning ("thinking") has demonstrated substantial gains on verifiable math and coding tasks. To investigate the effects of RL and thinking in a CXR VLM, we perform large-scale SFT on CXR data to build an updated RadVLM based on Qwen3-VL, followed by a cold-start SFT stage that equips the model with basic thinking ability. We then apply Group Relative Policy Optimization (GRPO) with clinically grounded, task-specific rewards for report generation and visual grounding, and run matched RL experiments on both domain-specific and general-domain Qwen3-VL variants, with and without thinking. Across these settings, we find that while strong SFT remains crucial for high base performance, RL provides additional gains on both tasks, whereas explicit thinking does not appear to further improve results. Under a unified evaluation pipeline, the RL-optimized RadVLM models outperform their baseline counterparts and reach state-of-the-art performance on both report generation and grounding, highlighting clinically aligned RL as a powerful complement to SFT for medical VLMs.

Daphné Chopard, Jorge da Silva Gonçalves, Irene Cannistraci et al.

Selecting compact and informative gene subsets from single-cell transcriptomic data is essential for biomarker discovery, improving interpretability, and cost-effective profiling. However, most existing feature selection approaches either operate as multi-stage pipelines or rely on post hoc feature attribution, making selection and prediction weakly coupled. In this work, we present YOTO (you only train once), an end-to-end framework that jointly identifies discrete gene subsets and performs prediction within a single differentiable architecture. In our model, the prediction task directly guides which genes are selected, while the learned subsets, in turn, shape the predictive representation. This closed feedback loop enables the model to iteratively refine both what it selects and how it predicts during training. Unlike existing approaches, YOTO enforces sparsity so that only the selected genes contribute to inference, eliminating the need to train additional downstream classifiers. Through a multi-task learning design, the model learns shared representations across related objectives, allowing partially labeled datasets to inform one another, and discovering gene subsets that generalize across tasks without additional training steps. We evaluate YOTO on two representative single-cell RNA-seq datasets, showing that it consistently outperforms state-of-the-art baselines. These results demonstrate that sparse, end-to-end, multi-task gene subset selection improves predictive performance and yields compact and meaningful gene subsets, advancing biomarker discovery and single-cell analysis.

Claudio Fanconi, Moritz Vandenhirtz, Severin Husmann et al.

Prototype learning, a popular machine learning method designed for inherently interpretable decisions, leverages similarities to learned prototypes for classifying new data. While it is mainly applied in computer vision, in this work, we build upon prior research and further explore the extension of prototypical networks to natural language processing. We introduce a learned weighted similarity measure that enhances the similarity computation by focusing on informative dimensions of pre-trained sentence embeddings. Additionally, we propose a post-hoc explainability mechanism that extracts prediction-relevant words from both the prototype and input sentences. Finally, we empirically demonstrate that our proposed method not only improves predictive performance on the AG News and RT Polarity datasets over a previous prototype-based approach, but also improves the faithfulness of explanations compared to rationale-based recurrent convolutions.

Jonas Mirlach, Sonia Laguna, Julia E. Vogt

Machine unlearning aims to remove the influence of specific data from trained models while preserving general utility. Existing approximate unlearning methods often rely on performance-degradation heuristics, such as loss maximization or random labeling. However, these signals can be poorly conditioned, leading to unstable optimization and harming the model's generalization. We argue that unlearning should instead prioritize distributional indistinguishability, aligning the model's behavior on forget data with its behavior on truly unseen data. Motivated by this, we propose Reference-Guided Unlearning (ReGUn), a framework that leverages a disjoint held-out dataset to provide a principled, class-conditioned reference for distillation. We demonstrate across various model architectures, natural image datasets, and varying forget fractions that ReGUn consistently outperforms standard approximate baselines, achieving a superior forgetting-utility trade-off.

Sonia Laguna, Jorge da Silva Goncalves, Moritz Vandenhirtz et al.

Machine unlearning is rapidly becoming a practical requirement, driven by privacy regulations, data errors, and the need to remove harmful or corrupted training samples. Despite this, most existing methods tackle the problem purely from a post-hoc perspective. They attempt to erase the influence of targeted training samples through parameter updates that typically require access to the full training data. This creates a mismatch with real deployment scenarios where unlearning requests can be anticipated, revealing a fundamental limitation of post-hoc approaches. We propose \textit{unlearning by design}, a novel paradigm in which models are directly trained to support forgetting as an inherent capability. We instantiate this idea with Machine UNlearning via KEY deletion (MUNKEY), a memory augmented transformer that decouples instance-specific memorization from model weights. Here, unlearning corresponds to removing the instance-identifying key, enabling direct zero-shot forgetting without weight updates or access to the original samples or labels. Across natural image benchmarks, fine-grained recognition, and medical datasets, MUNKEY outperforms all post-hoc baselines. Our results establish that unlearning by design enables fast, deployment-oriented unlearning while preserving predictive performance.

Nicolas Dickenmann, Yanis Merzouki, Sonia Laguna et al.

EasyRead pictograms are simple, visually clear images that represent specific concepts and support comprehension for people with intellectual disabilities, low literacy, or language barriers. The large-scale production of EasyRead content has traditionally been constrained by the cost and expertise required to manually design pictograms. In contrast, automatic generation of such images could significantly reduce production time and cost, enabling broader accessibility across digital and printed materials. However, modern diffusion-based image generation models tend to produce outputs that exhibit excessive visual detail and lack stylistic stability across random seeds, limiting their suitability for clear and consistent pictogram generation. This challenge highlights the need for methods specifically tailored to accessibility-oriented visual content. In this work, we present a unified pipeline for generating EasyRead pictograms by fine-tuning a Stable Diffusion model using LoRA adapters on a curated corpus that combines augmented samples from multiple pictogram datasets. Since EasyRead pictograms lack a unified formal definition, we introduce an EasyRead score to benchmark pictogram quality and consistency. Our results demonstrate that diffusion models can be effectively steered toward producing coherent EasyRead-style images, indicating that generative models can serve as practical tools for scalable and accessible pictogram production.

Patrik Reizinger, Alice Bizeul, Attila Juhos et al.

Supervised learning has become a cornerstone of modern machine learning, yet a comprehensive theory explaining its effectiveness remains elusive. Empirical phenomena, such as neural analogy-making and the linear representation hypothesis, suggest that supervised models can learn interpretable factors of variation in a linear fashion. Recent advances in self-supervised learning, particularly nonlinear Independent Component Analysis, have shown that these methods can recover latent structures by inverting the data generating process. We extend these identifiability results to parametric instance discrimination, then show how insights transfer to the ubiquitous setting of supervised learning with cross-entropy minimization. We prove that even in standard classification tasks, models learn representations of ground-truth factors of variation up to a linear transformation. We corroborate our theoretical contribution with a series of empirical studies. First, using simulated data matching our theoretical assumptions, we demonstrate successful disentanglement of latent factors. Second, we show that on DisLib, a widely-used disentanglement benchmark, simple classification tasks recover latent structures up to linear transformations. Finally, we reveal that models trained on ImageNet encode representations that permit linear decoding of proxy factors of variation. Together, our theoretical findings and experiments offer a compelling explanation for recent observations of linear representations, such as superposition in neural networks. This work takes a significant step toward a cohesive theory that accounts for the unreasonable effectiveness of supervised deep learning.

Thomas M. Sutter, Yang Meng, Andrea Agostini et al.

Variational Autoencoders for multimodal data hold promise for many tasks in data analysis, such as representation learning, conditional generation, and imputation. Current architectures either share the encoder output, decoder input, or both across modalities to learn a shared representation. Such architectures impose hard constraints on the model. In this work, we show that a better latent representation can be obtained by replacing these hard constraints with a soft constraint. We propose a new mixture-of-experts prior, softly guiding each modality's latent representation towards a shared aggregate posterior. This approach results in a superior latent representation and allows each encoding to preserve information better from its uncompressed original features. In extensive experiments on multiple benchmark datasets and two challenging real-world datasets, we show improved learned latent representations and imputation of missing data modalities compared to existing methods.

Nicolas Deperrois, Hidetoshi Matsuo, Samuel Ruipérez-Campillo et al.

The widespread use of chest X-rays (CXRs), coupled with a shortage of radiologists, has driven growing interest in automated CXR analysis and AI-assisted reporting. While existing vision-language models (VLMs) show promise in specific tasks such as report generation or abnormality detection, they often lack support for interactive diagnostic capabilities. In this work we present RadVLM, a compact, multitask conversational foundation model designed for CXR interpretation. To this end, we curate a large-scale instruction dataset comprising over 1 million image-instruction pairs containing both single-turn tasks -- such as report generation, abnormality classification, and visual grounding -- and multi-turn, multi-task conversational interactions. After fine-tuning RadVLM on this instruction dataset, we evaluate it across different tasks along with re-implemented baseline VLMs. Our results show that RadVLM achieves state-of-the-art performance in conversational capabilities and visual grounding while remaining competitive in other radiology tasks. Ablation studies further highlight the benefit of joint training across multiple tasks, particularly for scenarios with limited annotated data. Together, these findings highlight the potential of RadVLM as a clinically relevant AI assistant, providing structured CXR interpretation and conversational capabilities to support more effective and accessible diagnostic workflows.

Samuel Ruipérez-Campillo, Alain Ryser, Thomas M. Sutter et al.

In the field of cardiac electrophysiology (EP), effectively reducing noise in intra-cardiac signals is crucial for the accurate diagnosis and treatment of arrhythmias and cardiomyopathies. However, traditional noise reduction techniques fall short in addressing the diverse noise patterns from various sources, often non-linear and non-stationary, present in these signals. This work introduces a Variational Autoencoder (VAE) model, aimed at improving the quality of intra-ventricular monophasic action potential (MAP) signal recordings. By constructing representations of clean signals from a dataset of 5706 time series from 42 patients diagnosed with ischemic cardiomyopathy, our approach demonstrates superior denoising performance when compared to conventional filtering methods commonly employed in clinical settings. We assess the effectiveness of our VAE model using various metrics, indicating its superior capability to denoise signals across different noise types, including time-varying non-linear noise frequently found in clinical settings. These results reveal that VAEs can eliminate diverse sources of noise in single beats, outperforming state-of-the-art denoising techniques and potentially improving treatment efficacy in cardiac EP.

Alice Bizeul, Thomas Sutter, Alain Ryser et al. · eth-zurich

Predicting masked from visible parts of an image is a powerful self-supervised approach for visual representation learning. However, the common practice of masking random patches of pixels exhibits certain failure modes, which can prevent learning meaningful high-level features, as required for downstream tasks. We propose an alternative masking strategy that operates on a suitable transformation of the data rather than on the raw pixels. Specifically, we perform principal component analysis and then randomly mask a subset of components, which accounts for a fixed ratio of the data variance. The learning task then amounts to reconstructing the masked components from the visible ones. Compared to local patches of pixels, the principal components of images carry more global information. We thus posit that predicting masked from visible components involves more high-level features, allowing our masking strategy to extract more useful representations. This is corroborated by our empirical findings which demonstrate improved image classification performance for component over pixel masking. Our method thus constitutes a simple and robust data-driven alternative to traditional masked image modeling approaches.

Andrea Agostini, Daphné Chopard, Yang Meng et al.

Multimodal data integration and label scarcity pose significant challenges for machine learning in medical settings. To address these issues, we conduct an in-depth evaluation of the newly proposed Multimodal Variational Mixture-of-Experts (MMVM) VAE on the challenging MIMIC-CXR dataset. Our analysis demonstrates that the MMVM VAE consistently outperforms other multimodal VAEs and fully supervised approaches, highlighting its strong potential for real-world medical applications.

Andrea Agostini, Sonia Laguna, Alain Ryser et al.

Building generalizable medical AI systems requires pretraining strategies that are data-efficient and domain-aware. Unlike internet-scale corpora, clinical datasets such as MIMIC-CXR offer limited image counts and scarce annotations, but exhibit rich internal structure through multi-view imaging. We propose a self-supervised framework that leverages the inherent structure of medical datasets. Specifically, we treat paired chest X-rays (i.e., frontal and lateral views) as natural positive pairs, learning to reconstruct each view from sparse patches while aligning their latent embeddings. Our method requires no textual supervision and produces informative representations. Evaluated on MIMIC-CXR, we show strong performance compared to supervised objectives and baselines being trained without leveraging structure. This work provides a lightweight, modality-agnostic blueprint for domain-specific pretraining where data is structured but scarce

Boqi Chen, Cédric Vincent-Cuaz, Lydia A. Schoenpflug et al.

Vision foundation models (FMs) are accelerating the development of digital pathology algorithms and transforming biomedical research. These models learn, in a self-supervised manner, to represent histological features in highly heterogeneous tiles extracted from whole-slide images (WSIs) of real-world patient samples. The performance of these FMs is significantly influenced by the size, diversity, and balance of the pre-training data. However, data selection has been primarily guided by expert knowledge at the WSI level, focusing on factors such as disease classification and tissue types, while largely overlooking the granular details available at the tile level. In this paper, we investigate the potential of unsupervised automatic data curation at the tile-level, taking into account 350 million tiles. Specifically, we apply hierarchical clustering trees to pre-extracted tile embeddings, allowing us to sample balanced datasets uniformly across the embedding space of the pretrained FM. We further identify these datasets are subject to a trade-off between size and balance, potentially compromising the quality of representations learned by FMs, and propose tailored batch sampling strategies to mitigate this effect. We demonstrate the effectiveness of our method through improved performance on a diverse range of clinically relevant downstream tasks.

Alba Carballo-Castro, Sonia Laguna, Moritz Vandenhirtz et al.

Concept-based machine learning methods have increasingly gained importance due to the growing interest in making neural networks interpretable. However, concept annotations are generally challenging to obtain, making it crucial to leverage all their prior knowledge. By creating concept-enriched models that incorporate concept information into existing architectures, we exploit their interpretable capabilities to the fullest extent. In particular, we propose Concept-Guided Conditional Diffusion, which can generate visual representations of concepts, and Concept-Guided Prototype Networks, which can create a concept prototype dataset and leverage it to perform interpretable concept prediction. These results open up new lines of research by exploiting pre-existing information in the quest for rendering machine learning more human-understandable.

Sonia Laguna, Andrea Agostini, Alain Ryser et al.

Building robust medical machine learning systems requires pretraining strategies that exploit the intrinsic structure present in clinical data. We introduce Multiview Masked Autoencoder (MVMAE), a self-supervised framework that leverages the natural multi-view organization of radiology studies to learn view-invariant and disease-relevant representations. MVMAE combines masked image reconstruction with cross-view alignment, transforming clinical redundancy across projections into a powerful self-supervisory signal. We further extend this approach with MVMAE-V2T, which incorporates radiology reports as an auxiliary text-based learning signal to enhance semantic grounding while preserving fully vision-based inference. Evaluated on a downstream disease classification task on three large-scale public datasets, MIMIC-CXR, CheXpert, and PadChest, MVMAE consistently outperforms supervised and vision-language baselines. Furthermore, MVMAE-V2T provides additional gains, particularly in low-label regimes where structured textual supervision is most beneficial. Together, these results establish the importance of structural and textual supervision as complementary paths toward scalable, clinically grounded medical foundation models.

Wiktor Jan Hoffmann, Sonia Laguna, Moritz Vandenhirtz et al.

Concept Bottleneck Models (CBMs) are interpretable models that predict the target variable through high-level human-understandable concepts, allowing users to intervene on mispredicted concepts to adjust the final output. While recent work has shown that modeling dependencies between concepts can improve CBM performance, especially under interventions, such approaches typically require retraining the entire model, which may be infeasible when access to the original data or compute is limited. In this paper, we introduce Post-hoc Stochastic Concept Bottleneck Models (PSCBMs), a lightweight method that augments any pre-trained CBM with a multivariate normal distribution over concepts by adding only a small covariance-prediction module, without retraining the backbone model. We propose two training strategies and show on real-world data that PSCBMs consistently match or improve both concept and target accuracy over standard CBMs at test time. Furthermore, we show that due to the modeling of concept dependencies, PSCBMs perform much better than CBMs under interventions, while remaining far more efficient than retraining a similar stochastic model from scratch.

Yves Stebler, Thomas M. Sutter, Ece Ozkan et al.

Ultrasound (US) imaging is a critical tool in medical diagnostics, offering real-time visualization of physiological processes. One of its major advantages is its ability to capture temporal dynamics, which is essential for assessing motion patterns in applications such as cardiac monitoring, fetal development, and vascular imaging. Despite its importance, current deep learning models often overlook the temporal continuity of ultrasound sequences, analyzing frames independently and missing key temporal dependencies. To address this gap, we propose a method for learning effective temporal representations from ultrasound videos, with a focus on echocardiography-based ejection fraction (EF) estimation. EF prediction serves as an ideal case study to demonstrate the necessity of temporal learning, as it requires capturing the rhythmic contraction and relaxation of the heart. Our approach leverages temporally consistent masking and contrastive learning to enforce temporal coherence across video frames, enhancing the model's ability to represent motion patterns. Evaluated on the EchoNet-Dynamic dataset, our method achieves a substantial improvement in EF prediction accuracy, highlighting the importance of temporally-aware representation learning for real-time ultrasound analysis.

Max Krähenmann, Sergio Tascon-Morales, Fabian Laumer et al.

Volumetric ultrasound has the potential to significantly improve diagnostic accuracy and clinical decision-making, yet its widespread adoption remains limited by dependence on specialized hardware and restrictive acquisition protocols. In this work, we present a novel unsupervised framework for reconstructing 3D anatomical structures from freehand 2D transvaginal ultrasound (TVS) sweeps, without requiring external tracking or learned pose estimators. Our method adapts the principles of Gaussian Splatting to the domain of ultrasound, introducing a slice-aware, differentiable rasterizer tailored to the unique physics and geometry of ultrasound imaging. We model anatomy as a collection of anisotropic 3D Gaussians and optimize their parameters directly from image-level supervision, leveraging sensorless probe motion estimation and domain-specific geometric priors. The result is a compact, flexible, and memory-efficient volumetric representation that captures anatomical detail with high spatial fidelity. This work demonstrates that accurate 3D reconstruction from 2D ultrasound images can be achieved through purely computational means, offering a scalable alternative to conventional 3D systems and enabling new opportunities for AI-assisted analysis and diagnosis.

Lucas Erlacher, Samuel Ruipérez-Campillo, Holger Michel et al.

Pulmonary hypertension (PH) in newborns is a critical condition characterized by elevated pressure in the pulmonary arteries, leading to right ventricular strain and heart failure. While right heart catheterization (RHC) is the diagnostic gold standard, echocardiography is preferred due to its non-invasive nature, safety, and accessibility. However, its accuracy highly depends on the operator, making PH assessment subjective. While automated detection methods have been explored, most models focus on adults and rely on single-view echocardiographic frames, limiting their performance in diagnosing PH in newborns. While multi-view echocardiography has shown promise in improving PH assessment, existing models struggle with generalizability. In this work, we employ a multi-view variational autoencoder (VAE) for PH prediction using echocardiographic videos. By leveraging the VAE framework, our model captures complex latent representations, improving feature extraction and robustness. We compare its performance against single-view and supervised learning approaches. Our results show improved generalization and classification accuracy, highlighting the effectiveness of multi-view learning for robust PH assessment in newborns.

Sonia Laguna, Katarzyna Kobalczyk, Julia E. Vogt et al.

We introduce Concept Bottleneck Reward Models (CB-RM), a reward modeling framework that enables interpretable preference learning through selective concept annotation. Unlike standard RLHF methods that rely on opaque reward functions, CB-RM decomposes reward prediction into human-interpretable concepts. To make this framework efficient in low-supervision settings, we formalize an active learning strategy that dynamically acquires the most informative concept labels. We propose an acquisition function based on Expected Information Gain and show that it significantly accelerates concept learning without compromising preference accuracy. Evaluated on the UltraFeedback dataset, our method outperforms baselines in interpretability and sample efficiency, marking a step towards more transparent, auditable, and human-aligned reward models.

Nicola Bernold, Moritz Vandenhirtz, Alice Bizeul et al.

Text-to-image diffusion models generate images by iteratively denoising random noise, conditioned on a prompt. While these models have enabled impressive progress in image generation, they often fail to accurately reflect all semantic information described in the prompt -- failures that are difficult to detect automatically. In this work, we introduce a diffusion model architecture built with B-cos modules that offers inherent interpretability. Our approach provides insight into how individual prompt tokens affect the generated image by producing explanations that highlight the pixel regions influenced by each token. We demonstrate that B-cos diffusion models can produce high-quality images while providing meaningful insights into prompt-image alignment.

Moritz Vandenhirtz, Julia E. Vogt

Understanding the decision-making process of machine learning models provides valuable insights into the task, the data, and the reasons behind a model's failures. In this work, we propose a method that performs inherently interpretable predictions through the instance-wise sparsification of input images. To align the sparsification with human perception, we learn the masking in the space of semantically meaningful pixel regions rather than on pixel-level. Additionally, we introduce an explicit way to dynamically determine the required level of sparsity for each instance. We show empirically on semi-synthetic and natural image datasets that our inherently interpretable classifier produces more meaningful, human-understandable predictions than state-of-the-art benchmarks.

Daphné Chopard, Sonia Laguna, Kieran Chin-Cheong et al.

General movements (GMs) are spontaneous, coordinated body movements in infants that offer valuable insights into the developing nervous system. Assessed through the Prechtl GM Assessment (GMA), GMs are reliable predictors for neurodevelopmental disorders. However, GMA requires specifically trained clinicians, who are limited in number. To scale up newborn screening, there is a need for an algorithm that can automatically classify GMs from infant video recordings. This data poses challenges, including variability in recording length, device type, and setting, with each video coarsely annotated for overall movement quality. In this work, we introduce a tool for extracting features from these recordings and explore various machine learning techniques for automated GM classification.

Jorge da Silva Gonçalves, Laura Manduchi, Moritz Vandenhirtz et al.

Generative modeling and clustering are conventionally distinct tasks in machine learning. Variational Autoencoders (VAEs) have been widely explored for their ability to integrate both, providing a framework for generative clustering. However, while VAEs can learn meaningful cluster representations in latent space, they often struggle to generate high-quality samples. This paper addresses this problem by introducing TreeDiffusion, a deep generative model that conditions diffusion models on learned latent hierarchical cluster representations from a VAE to obtain high-quality, cluster-specific generations. Our approach consists of two steps: first, a VAE-based clustering model learns a hierarchical latent representation of the data. Second, a cluster-aware diffusion model generates realistic images conditioned on the learned hierarchical structure. We systematically compare the generative capabilities of our approach with those of alternative conditioning strategies. Empirically, we demonstrate that conditioning diffusion models on hierarchical cluster representations improves the generative performance on real-world datasets compared to other approaches. Moreover, a key strength of our method lies in its ability to generate images that are both representative and specific to each cluster, enabling more detailed visualization of the learned latent structure. Our approach addresses the generative limitations of VAE-based clustering approaches by leveraging their learned structure, thereby advancing the field of generative clustering.

Moritz Vandenhirtz, Florian Barkmann, Laura Manduchi et al.

We propose a novel method, scTree, for single-cell Tree Variational Autoencoders, extending a hierarchical clustering approach to single-cell RNA sequencing data. scTree corrects for batch effects while simultaneously learning a tree-structured data representation. This VAE-based method allows for a more in-depth understanding of complex cellular landscapes independently of the biasing effects of batches. We show empirically on seven datasets that scTree discovers the underlying clusters of the data and the hierarchical relations between them, as well as outperforms established baseline methods across these datasets. Additionally, we analyze the learned hierarchy to understand its biological relevance, thus underpinning the importance of integrating batch correction directly into the clustering procedure.

Moritz Vandenhirtz, Sonia Laguna, Ričards Marcinkevičs et al.

Concept Bottleneck Models (CBMs) have emerged as a promising interpretable method whose final prediction is based on intermediate, human-understandable concepts rather than the raw input. Through time-consuming manual interventions, a user can correct wrongly predicted concept values to enhance the model's downstream performance. We propose Stochastic Concept Bottleneck Models (SCBMs), a novel approach that models concept dependencies. In SCBMs, a single-concept intervention affects all correlated concepts, thereby improving intervention effectiveness. Unlike previous approaches that model the concept relations via an autoregressive structure, we introduce an explicit, distributional parameterization that allows SCBMs to retain the CBMs' efficient training and inference procedure. Additionally, we leverage the parameterization to derive an effective intervention strategy based on the confidence region. We show empirically on synthetic tabular and natural image datasets that our approach improves intervention effectiveness significantly. Notably, we showcase the versatility and usability of SCBMs by examining a setting with CLIP-inferred concepts, alleviating the need for manual concept annotations.

Kacper Sokol, Julia E. Vogt

Despite significant progress, evaluation of explainable artificial intelligence remains elusive and challenging. In this paper we propose a fine-grained validation framework that is not overly reliant on any one facet of these sociotechnical systems, and that recognises their inherent modular structure: technical building blocks, user-facing explanatory artefacts and social communication protocols. While we concur that user studies are invaluable in assessing the quality and effectiveness of explanation presentation and delivery strategies from the explainees' perspective in a particular deployment context, the underlying explanation generation mechanisms require a separate, predominantly algorithmic validation strategy that accounts for the technical and human-centred desiderata of their (numerical) outputs. Such a comprehensive sociotechnical utility-based evaluation framework could allow to systematically reason about the properties and downstream influence of different building blocks from which explainable artificial intelligence systems are composed -- accounting for a diverse range of their engineering and social aspects -- in view of the anticipated use case.

Mike Laszkiewicz, Imant Daunhawer, Julia E. Vogt et al.

Recent years have witnessed a rapid development of deep generative models for creating synthetic media, such as images and videos. While the practical applications of these models in everyday tasks are enticing, it is crucial to assess the inherent risks regarding their fairness. In this work, we introduce a comprehensive framework for benchmarking the performance and fairness of conditional generative models. We develop a set of metrics$\unicode{x2013}$inspired by their supervised fairness counterparts$\unicode{x2013}$to evaluate the models on their fairness and diversity. Focusing on the specific application of image upsampling, we create a benchmark covering a wide variety of modern upsampling methods. As part of the benchmark, we introduce UnfairFace, a subset of FairFace that replicates the racial distribution of common large-scale face datasets. Our empirical study highlights the importance of using an unbiased training set and reveals variations in how the algorithms respond to dataset imbalances. Alarmingly, we find that none of the considered methods produces statistically fair and diverse results. All experiments can be reproduced using our provided repository.

Sonia Laguna, Ričards Marcinkevičs, Moritz Vandenhirtz et al.

Recently, interpretable machine learning has re-explored concept bottleneck models (CBM). An advantage of this model class is the user's ability to intervene on predicted concept values, affecting the downstream output. In this work, we introduce a method to perform such concept-based interventions on pretrained neural networks, which are not interpretable by design, only given a small validation set with concept labels. Furthermore, we formalise the notion of intervenability as a measure of the effectiveness of concept-based interventions and leverage this definition to fine-tune black boxes. Empirically, we explore the intervenability of black-box classifiers on synthetic tabular and natural image benchmarks. We focus on backbone architectures of varying complexity, from simple, fully connected neural nets to Stable Diffusion. We demonstrate that the proposed fine-tuning improves intervention effectiveness and often yields better-calibrated predictions. To showcase the practical utility of our techniques, we apply them to deep chest X-ray classifiers and show that fine-tuned black boxes are more intervenable than CBMs. Lastly, we establish that our methods are still effective under vision-language-model-based concept annotations, alleviating the need for a human-annotated validation set.

Moritz Vandenhirtz, Laura Manduchi, Ričards Marcinkevičs et al.

Spurious correlations are everywhere. While humans often do not perceive them, neural networks are notorious for learning unwanted associations, also known as biases, instead of the underlying decision rule. As a result, practitioners are often unaware of the biased decision-making of their classifiers. Such a biased model based on spurious correlations might not generalize to unobserved data, leading to unintended, adverse consequences. We propose Signal is Harder (SiH), a variational-autoencoder-based method that simultaneously trains a biased and unbiased classifier using a novel, disentangling reweighting scheme inspired by the focal loss. Using the unbiased classifier, SiH matches or improves upon the performance of state-of-the-art debiasing methods. To improve the interpretability of our technique, we propose a perturbation scheme in the latent space for visualizing the bias that helps practitioners become aware of the sources of spurious correlations.

Thomas M. Sutter, Alain Ryser, Joram Liebeskind et al.

Partitioning a set of elements into an unknown number of mutually exclusive subsets is essential in many machine learning problems. However, assigning elements, such as samples in a dataset or neurons in a network layer, to an unknown and discrete number of subsets is inherently non-differentiable, prohibiting end-to-end gradient-based optimization of parameters. We overcome this limitation by proposing a novel two-step method for inferring partitions, which allows its usage in variational inference tasks. This new approach enables reparameterized gradients with respect to the parameters of the new random partition model. Our method works by inferring the number of elements per subset and, second, by filling these subsets in a learned order. We highlight the versatility of our general-purpose approach on three different challenging experiments: variational clustering, inference of shared and independent generative factors under weak supervision, and multitask learning.

Imant Daunhawer, Thomas M. Sutter, Kieran Chin-Cheong et al.

Multimodal variational autoencoders (VAEs) have shown promise as efficient generative models for weakly-supervised data. Yet, despite their advantage of weak supervision, they exhibit a gap in generative quality compared to unimodal VAEs, which are completely unsupervised. In an attempt to explain this gap, we uncover a fundamental limitation that applies to a large family of mixture-based multimodal VAEs. We prove that the sub-sampling of modalities enforces an undesirable upper bound on the multimodal ELBO and thereby limits the generative quality of the respective models. Empirically, we showcase the generative quality gap on both synthetic and real data and present the tradeoffs between different variants of multimodal VAEs. We find that none of the existing approaches fulfills all desired criteria of an effective multimodal generative model when applied on more complex datasets than those used in previous benchmarks. In summary, we identify, formalize, and validate fundamental limitations of VAE-based approaches for modeling weakly-supervised data and discuss implications for real-world applications.

Laura Manduchi, Kieran Chin-Cheong, Holger Michel et al.

Constrained clustering has gained significant attention in the field of machine learning as it can leverage prior information on a growing amount of only partially labeled data. Following recent advances in deep generative models, we propose a novel framework for constrained clustering that is intuitive, interpretable, and can be trained efficiently in the framework of stochastic gradient variational inference. By explicitly integrating domain knowledge in the form of probabilistic relations, our proposed model (DC-GMM) uncovers the underlying distribution of data conditioned on prior clustering preferences, expressed as pairwise constraints. These constraints guide the clustering process towards a desirable partition of the data by indicating which samples should or should not belong to the same cluster. We provide extensive experiments to demonstrate that DC-GMM shows superior clustering performances and robustness compared to state-of-the-art deep constrained clustering methods on a wide range of data sets. We further demonstrate the usefulness of our approach on two challenging real-world applications.