Publications

A Layer Selection Approach to Test Time Adaptation

Sabyasachi Sahoo

Mostafa Elaraby

Jonas Ngnawe

Yann Batiste Pequignot

Frédéric Precioso

Christian Gagné

Test Time Adaptation (TTA) addresses the problem of distribution shift by adapting a pretrained model to a new domain during inference. When… (voir plus) faced with challenging shifts, most methods collapse and perform worse than the original pretrained model. In this paper, we find that not all layers are equally receptive to the adaptation, and the layers with the most misaligned gradients often cause performance degradation. To address this, we propose GALA, a novel layer selection criterion to identify the most beneficial updates to perform during test time adaptation. This criterion can also filter out unreliable samples with noisy gradients. Its simplicity allows seamless integration with existing TTA loss functions, thereby preventing degradation and focusing adaptation on the most trainable layers. This approach also helps to regularize adaptation to preserve the pretrained features, which are crucial for handling unseen domains. Through extensive experiments, we demonstrate that the proposed layer selection framework improves the performance of existing TTA approaches across multiple datasets, domain shifts, model architectures, and TTA losses.

2025-04-10

Proceedings of the AAAI Conference on Artificial Intelligence (publié)

openreview.net

StarVector: Generating Scalable Vector Graphics Code from Images and Text

Juan A. Rodriguez

Abhay Puri

Shubham Agarwal

Issam H. Laradji

Juan A. Rodriguez

Sai Rajeswar

David Vázquez

Christopher Pal

Marco Pedersoli

Scalable Vector Graphics (SVGs) are vital for modern image rendering due to their scalability and versatility. Previous SVG generation metho… (voir plus)ds have focused on curve-based vectorization, lacking semantic understanding, often producing artifacts, and struggling with SVG primitives beyond path curves. To address these issues, we introduce StarVector, a multimodal large language model for SVG generation. It performs image vectorization by understanding image semantics and using SVG primitives for compact, precise outputs. Unlike traditional methods, StarVector works directly in the SVG code space, leveraging visual understanding to apply accurate SVG primitives. To train StarVector, we create SVG-Stack, a diverse dataset of 2M samples that enables generalization across vectorization tasks and precise use of primitives like ellipses, polygons, and text. We address challenges in SVG evaluation, showing that pixel-based metrics like MSE fail to capture the unique qualities of vector graphics. We introduce SVG-Bench, a benchmark across 10 datasets, and 3 tasks: Image-to-SVG, Text-to-SVG generation, and diagram generation. Using this setup, StarVector achieves state-of-the-art performance, producing more compact and semantically rich SVGs.

2025-04-10

Proceedings of the AAAI Conference on Artificial Intelligence (publié)

Genetic modulation of brain dynamics in neurodevelopmental disorders: the impact of copy number variations on resting-state EEG

Adrien E. E. Dubois

Elisabeth Audet-Duchesne

Inga Sophia Knoth

Charles-Olivier Martin

Khadije Jizi

Petra Tamer

Nadine Younis

Sébastien Jacquemont

Guillaume Dumas

Sarah Lippé

Research has shown that many copy number variations (CNVs) increase the risk of neurodevelopmental disorders (e.g., autism, ADHD, schizophre… (voir plus)nia). However, little is known about the effects of CNVs on brain development and function. Resting-state electroencephalography (EEG) is a suitable method to study the disturbances of neuronal functioning in CNVs. We aimed to determine whether there are resting-state EEG signatures that are characteristic of children with pathogenic CNVs. EEG resting-state brain activity of 109 CNV carriers (66 deletion carriers, 43 duplication carriers) aged 3 to 17 years was recorded for 4 minutes. To better account for developmental variations, EEG indices (power spectral density and functional connectivity) were corrected with a normative model estimated from 256 Healthy Brain Network controls. Results showed a decreased exponent of the aperiodic activity and a reduced alpha peak frequency in CNV carriers. Additionally, the study showed altered periodic components and connectivity in several frequency bands. Deletion and duplication carriers exhibited a similar overall pattern of deviations in spectral and connectivity measures, although the significance and effect sizes relative to the control group varied across frequency bands. Deletion and duplication carriers can be differentiated by their periodic power in the gamma band and connectivity in the low alpha band, with duplication carriers showing more disrupted alterations than deletion carriers. The distinctive alterations in spectral patterns were found to be most prominent during adolescence. The results suggest that CNV carriers show electrophysiological alterations compared to neurotypical controls, regardless of the gene dosage effect and their affected genomic region. At the same time, while duplications and deletions share common electrophysiological alterations, each exhibits distinct brain alteration signatures that reflect gene dosage-specific effects.

2025-04-09

Translational Psychiatry (publié)

Echoes in the Noise: Posterior Samples of Faint Galaxy Surface Brightness Profiles with Score-based Likelihoods and Priors

Connor Bottrell

Laurence Perreaul-Levasseur

Examining the detailed structure of galaxy populations provides valuable insights into their formation and evolution mechanisms. Significant… (voir plus) barriers to such analysis are the nontrivial noise properties of real astronomical images and the point-spread function, which blurs structure. Here we present a framework which combines recent advances in score-based likelihood characterization and diffusion model priors to perform a Bayesian analysis of image deconvolution. The method, when applied to minimally processed Hubble Space Telescope data, recovers structures which have otherwise only become visible in next-generation James Webb Space Telescope imaging.

2025-04-07

The Astronomical Journal (publié)

InfoGain Wavelets: Furthering the Design of Diffusion Wavelets for Graph-Structured Data

David R. Johnson

Smita Krishnaswamy

Michael Perlmutter

Diffusion wavelets extract information from graph signals at different scales of resolution by utilizing graph diffusion operators raised to… (voir plus) various powers, known as diffusion scales. Traditionally, the diffusion scales are chosen to be dyadic integers,

2025-04-07

ArXiv (prépublication)

Advancing Sustainable Maritime Transport: A Machine Learning Approach to Predict and Mitigate Underwater Radiated Noise from Ships

Soukaina Boujdi

Ayoub Atanane

Pierre Cauchy

2025-04-06

Proceedings of the International Conference on Industrial Engineering and Operations Management (publié)

A Comparative Analysis of AI Models for Short-Term Solar Irradiance Forecasting

Saad Benbrahim

Abdelaziz Berrado

Hanane Dagdougui

2025-04-06

Proceedings of the International Conference on Industrial Engineering and Operations Management (publié)

ECLARE: multi-teacher contrastive learning via ensemble distillation for diagonal integration of single-cell multi-omic data

Dylan Mann-Krzisnik

Anjali Chawla

Gustavo Turecki

Corina Nagy

Yue Li

Integrating multimodal single-cell data such as scRNA-seq with scATAC-seq is essential for decoding gene regulatory networks, but remains di… (voir plus)fficult due to feature harmonization and limited paired multiome data. We introduce ECLARE, a framework that uses multi-teacher ensemble knowledge distillation with contrastive learning and optimal-transport alignment to integrate unpaired single-cell multi-omic datasets. Across benchmarks, ECLARE achieves competitive performance for multimodal integration and biological structure preservation. We further demonstrate utility in a major depressive disorder case study using unpaired snRNA-seq and snATAC-seq, identifying transcription factor–target gene programs that are differentially regulated with sex- and cell-type specificity. Finally, ECLARE learns continuous representations that capture longitudinal structure, highlighting altered neurodevelopmental programs associated with depression in female subjects. Altogether, ECLARE expands the practical reach of multimodal single-cell analysis by enabling diagonal integration of unpaired data with strong biological preservation, facilitating integrative regulatory studies across diverse cohorts and conditions.

2025-04-06

bioRxiv (prépublication)

Enhancing Hybrid Model for Photovoltaic Power Prediction: A Case Study of Morocco

Samira Abousaid

Abdelaziz Berrado

Hanane Dagdougui

2025-04-06

Proceedings of the International Conference on Industrial Engineering and Operations Management (publié)

Predicting greenhouse gas Emissions in Shipping: A Case Study Of Canada

Abdelhak El Aissi

Abdelaziz Berrado

Stephane Carron

2025-04-06

Proceedings of the International Conference on Industrial Engineering and Operations Management (publié)

DiffKillR: Killing and Recreating Diffeomorphisms for Cell Annotation in Dense Microscopy Images

Chen Liu

Danqi Liao

Alejandro Parada-Mayorga

Alejandro Ribeiro

Marcello DiStasio

Smita Krishnaswamy

The proliferation of digital microscopy images, driven by advances in automated whole slide scanning, presents significant opportunities for… (voir plus) biomedical research and clinical diagnostics. However, accurately annotating densely packed information in these images remains a major challenge. To address this, we introduce DiffKillR, a novel framework that reframes cell annotation as the combination of archetype matching and image registration tasks. DiffKillR employs two complementary neural networks: one that learns a diffeomorphism-invariant feature space for robust cell matching and another that computes the precise warping field between cells for annotation mapping. Using a small set of annotated archetypes, DiffKillR efficiently propagates annotations across large microscopy images, reducing the need for extensive manual labeling. More importantly, it is suitable for any type of pixel-level annotation. We will discuss the theoretical properties of DiffKillR and validate it on three microscopy tasks, demonstrating its advantages over existing supervised, semi-supervised, and unsupervised methods.

2025-04-05

ICASSP 2025 - 2025 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) (publié)