Portrait de Guy Wolf

Guy Wolf

Membre académique principal
Chaire en IA Canada-CIFAR
Professeur titulaire, Université de Montréal
Professeur affilié, Concordia University, Département d'informatique et de génie logiciel
Chief AI Scientist, DRI CHUM and Associate Director, AI & Data Science, CRCHUM and Scientific Director, CITADEL, CR/CHUM and Principal Scientist, CRCHUM
Sujets de recherche
Apprentissage automatique médical
Apprentissage de représentations
Apprentissage multimodal
Apprentissage profond
Apprentissage spectral
Apprentissage sur graphes
Exploration des données
Modélisation moléculaire
Recherche d'information
Réseaux de neurones en graphes
Systèmes dynamiques
Théorie de l'apprentissage automatique

Biographie

Guy Wolf est scientifique en chef en IA au Centre hospitalier de l’Université de Montréal (CHUM), directeur associé en IA et science des données de son centre de recherche (CRCHUM), ainsi que directeur scientifique de son centre d'intégration et d'analyse des données médicales (CITADEL). Il est professeur titulaire à l'Université de Montréal (UdeM) et membre académique principal à Mila, où il est titulaire d'une chaire en IA Canada-CIFAR. Il est également professeur IVADO, membre du Centre de recherches mathématiques (CRM) et chercheur principal participant au sein du laboratoire international Helmholtz sur la dynamique causale des cellules.

En 2024, il a obtenu une bourse de recherche Humboldt pour chercheurs confirmés, dans le cadre de laquelle il a été professeur invité à l'Université de Heidelberg (2024) et à Helmholtz Munich (2024-2026), en Allemagne. Avant de se joindre à l'UdeM et à Mila, il a été professeur adjoint Gibbs (2015-2018) et chercheur associé (2018) à l'Université Yale (Connecticut, États-Unis), ainsi que chercheur postdoctoral (2013-2015) à l'École normale supérieure de Paris (France). Il est titulaire d'un doctorat en informatique de l'Université de Tel-Aviv (Israël) et compte cinq années d'expérience préalable en conception et développement de logiciels d'analyse de données en milieu militaire.

Ses travaux actuels portent sur l'apprentissage guidé de représentations pour l'exploration de données, notamment à l'aide de méthodes qui exploitent l'apprentissage sur variétés (manifold learning) et l'apprentissage profond géométrique pour la réduction de dimensionnalité, la visualisation, le débruitage, l'augmentation de données et le changement d'échelle (coarse-graining). Bien que pertinentes pour une vaste gamme d'applications, ses recherches s'intéressent tout particulièrement au croisement de l'IA et de la santé, notamment aux outils d'analyse exploratoire des données biomédicales en multiomique unicellulaire, en découverte de médicaments et en neurosciences.

Étudiants actuels

Collaborateur·rice de recherche - Yale University
Co-superviseur⋅e :
Collaborateur·rice de recherche - University of Tübingen
Maîtrise recherche - UdeM
Co-superviseur⋅e :
Maîtrise recherche - Concordia
Superviseur⋅e principal⋅e :
Doctorat - Concordia
Superviseur⋅e principal⋅e :
Visiteur de recherche indépendant - Helmholtz Munich
Doctorat - UdeM
Co-superviseur⋅e :
Maîtrise recherche - Concordia
Superviseur⋅e principal⋅e :
Collaborateur·rice de recherche
Postdoctorat - Concordia
Superviseur⋅e principal⋅e :
Doctorat - Concordia
Superviseur⋅e principal⋅e :
Collaborateur·rice de recherche - BYU
Visiteur de recherche indépendant - University of Fribourg
Doctorat - UdeM
Superviseur⋅e principal⋅e :
Doctorat - Concordia
Superviseur⋅e principal⋅e :
Maîtrise recherche - McGill
Collaborateur·rice alumni - UdeM
Co-superviseur⋅e :

Publications

Can Model Merging Improve Aggregation in DiLoCo?
Model merging techniques, which aggregate independently finetuned models into one to combine their capabilities, have become a topic of sign… (voir plus)ificant interest in recent years, with a broad array of methods having been proposed to tackle this problem. Simultaneously, an emerging trend in distributed learning has been the use of methods such as local SGD and DiLoCo, which greatly reduce communication costs by periodically aggregating the independently trained local models. However, these communication-efficient methods have been shown to degrade in performance relative to the FLOP-matched data-parallel gold standard as the number of independent local models grows and as the number of local training steps before global communication is increased. In this work, we draw an explicit analogy between the pseudo-gradient aggregation step in local SGD/DiLoCo and task arithmetic-based model merging, establishing a straightforward way to utilize merging methods in the context of distributed optimization. We then evaluate multiple state-of-the-art model merging methods in this setting and identify one method in particular, Iso-C, as a promising approach for improving DiLoCo. We find that DiLoCo SGD with Iso-C aggregation outperforms not only simple pseudo-gradient averaging but even the momentum-based DiLoCo, despite lacking a momentum mechanism itself. Building on this finding, we propose IsoLoCo, which adapts Iso-C for distributed training by equipping it with Nesterov momentum. Our empirical evaluations on language model pre-training across varying numbers of local workers show that IsoLoCo significantly outperforms DiLoCo, with the gap between them widening as the number of workers increases. This advantage remains present across model sizes and inner step counts, confirming that merging-inspired aggregation is an effective strategy for low-communication distributed training.
Path-independent Flow Matching for Multi-parameter Generative Dynamics
Flow Matching is a powerful framework for learning transport maps between probability distributions. Yet its standard single-parameter formu… (voir plus)lation is not designed to capture multi-parameter variations where the resulting transport should be path-independent. Path independence is crucial because it ensures that transformations depend only on the initial and target distributions, not on the specific path. In this work, we introduce Path-independent Flow Matching (PiFM), a method for learning vector fields whose induced flows yield path-independent transport between distributions. We show that PiFM generalizes Flow Matching to higher-dimensional parameter domains while enforcing structural conditions that ensure consistency of composed transformations. In addition, we show that, under suitable assumptions, PiFM approximates the Wasserstein barycenter, linking the framework to a notion of distributional interpolation. To enable practical training, we propose a tractable, simulation-free objective that regresses onto multi-parameter conditional probability paths. We showcase empirically that PiFM outperforms other approaches on both synthetic and real world data in interpolating path-independent trajectories and generating desired out of distribution samples.
PHATE-Derived Ancestry Coordinates Capture Continuous Population Structure in Human Genomic Data
Jean‐Christophe Grenier
Raphaël Poujol
Mohamed Elyes Kanoun
João Felipe Rocha
Kevin R. Moon
Abstract Population genetic data are increasingly visualized using low-dimensional embeddings as ancestry coordinates, yet how these methods… (voir plus) represent population structure can depend strongly on sampling and construction choices. Principal component analysis (PCA), when visualized using only leading components, can obscure fine-scale structure, while nonlinear neighbor-based methods such as uniform manifold approximation and projection (UMAP) often fragment geographic and admixture gradients into discrete clusters. Here, we introduce a framework for constructing ancestry coordinates using potential of heat-diffusion for affinity-based trajectory embedding (PHATE), a diffusion-based method that preserves relationships across scales. Across synthetic benchmarks and large genomic datasets, we show that appropriate scale tuning and population balancing allow PHATE to recover geographic organization and admixture gradients. We show that incorrectly tuned neighborhood scale and population imbalance can distort the resulting representations of population structure. These results therefore provide a methodological framework and practical guidance for constructing ancestry visualizations that better reflect continuous human genetic diversity.
GeoSinkhorn Flow: Geometry-Aware Flow Matching for Perturbation-Conditioned Dynamics in Single cell Data Phenoscapes
Ke Xu
Alistair Wilkinson
Ni Shuang
Gerstein Mark
Christopher J. Tape
Single-cell technologies increasingly enable large-scale profiling of complex biological systems across many experimental perturbations, but… (voir plus) understanding how a system reorganizes across an entire space of conditions is difficult to interpret from single-cell embeddings alone. Phenoscapes address this by embedding entire cell-state distributions, rather than individual cells, as points in a data-driven landscape. Thus, a single dot represents one experimental condition rather than one cell, yielding a map of the perturbation or patient landscape that reveals how phenotype varies continuously across conditions. Phenoscapes remain a comparatively new and underexplored representation and existing static phenoscapes position experimental conditions relative to one another without specifying trajectories that describe how a cellular population moves from one condition to another. Flow- and trajectory-based models for single-cell data are numerous, but existing approaches operate at the level of individual cells, are not built to respect the geometry of a phenoscape, and typically cannot generalize their predicted flows to conditions not seen during training. Here we introduce GeoSinkhorn Flow, a perturbation-conditioned, geometry-aware flow-matching framework specialized designed to convert a static phenoscape into a dynamic phenoscape, in which each condition-level point is additionally connected by geodesic, condition-aware transport trajectories that are consistent with the phenoscape's underlying geometry. GeoSinkhorn Flow can also model transitions directly between time points without constructing a phenoscape. GeoSinkhorn Flow uses heat-diffusion geometry to define manifold-aware transport costs and trains a geometric autoencoder whose latent Euclidean distances stay aligned with the static phenoscape geometry while extending it to unseen cells. By explicitly conditioning the learned transport dynamics on perturbation-specific gene or drug embeddings, GeoSinkhorn Flow generalizes its trajectories to new measurement conditions, produces intermediate progression states along a trajectory, and extends predictions to unseen cells and held-out perturbations. An unbalanced optimal transport variant further learns growth or transition weights, so that changes in relative cell-state abundance or transition propensity are modeled together with state movement. Across synthetic manifolds, patient-derived organoid drug responses, Perturb-seq knockdowns, and time-resolved CyTOF signaling, GeoSinkhorn Flow recovered intrinsic manifold geometry, inferred biologically coherent response trajectories, and generalized to held-out drug combinations, perturbation genes, and time points. Together, these results establish GeoSinkhorn Flow as a perturbation-conditioned, phenoscape-consistent dynamical framework, distinct from standard flow-matching approaches, for converting static phenoscapes into dynamic ones that generalize across experimental conditions.
Path Laplacian Encodings for Directed Graphs
Michael Perlmutter
Bastian Rieck
Directed graphs naturally model many real-world systems in which interactions are asymmetric, such as citation networks, web graphs, and inf… (voir plus)ormation-flow networks. However, graph learning methods commonly rely on message passing with symmetrized graph representations or positional encodings that only partially exploit edge directionality. We introduce PathLapPE, a novel spectral positional encoding (PE) derived from the path Laplacian on directed graphs. PathLapPE provides node- and edge-level features that encode directional higher-order structure and can be incorporated into standard graph learning architectures. Empirical results on node- and graph-level benchmark tasks show that PathLapPE yields consistent improvements across several architectures, especially when combined with direction-aware message passing. Compared with magnetic Laplacian positional encodings, a widely studied spectral positional encoding for directed graphs, PathLapPE does not require additional fine-tuning of directionality hyperparameters while offering competitive runtime and performance.
A Closed Form Depth-Width Law for Neural Collapse in Deep Linear Networks
Pankhil Gawade
Kirsten Fischer
Ewa Szczurek
How depth and width of neural networks shape learned representations is usually studied only after training. We show that, in a tractable fe… (voir plus)ature-learning setting, this dependence can instead be obtained in closed form. Using neural collapse as a probe of representation geometry, we express NC1--NC3 directly in terms of the learned feature kernel, extending an identity known for NC1 to the angular part of NC2 and to NC3 under a ridge readout. For deep linear networks with exchangeable class structure, the resulting kernel equations yield a geometric law, exact within the theory, for the layerwise evolution of within-class variability. At large width, the accumulated compression becomes a function of the depth-to-width ratio
Circuit Discovery Helps Detect LLM Jailbreaking: A Mechanistic Interpretability Study
Despite extensive safety alignment, large language models (LLMs) remain vulnerable to jailbreak attacks that bypass safeguards to elicit har… (voir plus)mful content. While prior work attributes this vulnerability to safety training limitations, the internal mechanisms by which LLMs process adversarial prompts remain poorly understood. We present a mechanistic analysis of the jailbreaking behavior in a large-scale, safety-aligned LLM, focusing on LLaMA-2-7B-chat-hf. Leveraging edge attribution patching and subnetwork probing, we systematically identify computational circuits responsible for generating affirmative responses to jailbreak prompts. Ablating these circuits during the first token prediction can reduce attack success rates by up to 80\%, demonstrating its critical role in safety bypass. Our analysis uncovers key attention heads and MLP pathways that mediate adversarial prompt exploitation, revealing how important tokens propagate through these components to override safety constraints. These findings advance the understanding of adversarial vulnerabilities in aligned LLMs and pave the way for targeted, interpretable defense mechanisms based on mechanistic interpretability.
Mapping Alzheimer’s neuropathology signatures to the whole brain transcriptome using machine learning data-fusion
L. M. Hodgson
David A Bennett
Elisabeth B. Binder
Abstract In Alzheimer’s disease (AD), misfolded proteins emerge across the entire brain in structured, yet not rigid, spatiotemporal patte… (voir plus)rns. Yet, a systematic bias of single-cell genomics toward sampling mostly cortical tissue limits our understanding of the whole-brain transcriptomic vulnerability to AD. Here, we develop a machine learning method to extrapolate local AD neuropathology signatures to the whole brain. By analyzing gene expression profiles of over two million cortical cells from 427 humans spanning the AD-pathology spectrum, we derive transcriptomic estimators of AD neuropathology. After extensive validations on datasets with known ground truth, we apply this framework to three million cells from 108 brain regions in the Siletti whole human brain atlas and derive an anticipated brain map of transcriptomic signatures indexing AD neuropathology. This interrogation of regions spanning the cortical, subcortical, and brainstem structures uncovers transcriptomic signatures associated with hyperphosphorylated tau in the medulla oblongata, dorsal raphe nucleus, and the tuberal and mammillary regions of the hypothalamus. At the cellular level, assessments of these signatures across 31 cell populations identify VGLUT1/2 expressing neurons, astrocytes, and microglia as key neuropathology-resembling populations. Within the hippocampus, pathology signatures surface in the rostral cornu ammonis (CA) subfields, particularly in the CA1 pyramidal neurons and dentate granule cells. β-amyloid-like signatures localize to the neocortex with laminar selectivity — most prominently in upper layer somatostatin+ intratelencephalic neurons (L2-L3), but also in deep layer intratelencephalic and corticothalamic neurons (L5-L6). Neocortical astrocytes and microglia exhibiting disease associated signatures similarly demonstrate a unique laminar preference. Together, this study provides the first whole human brain map of AD pathology-associated transcriptomic signals, and exposes cell type, region, and cortex layer specific vulnerabilities.
Freeze, Diffuse, Decode: Task-Aware Adaptation of Transformer Embeddings for Antimicrobial Peptide Design
Pankhil Gawade
Adam Izdebski
Kevin R. Moon
Jake Slater Rhodes
Ewa Szczurek
Pretrained transformers provide general-purpose molecular embeddings for downstream tasks. While these embeddings provide task-agnostic stru… (voir plus)ctural patterns, they lack task-specific alignment, limiting downstream performance. Here, we introduce Freeze, Diffuse, Decode (FDD), a diffusion-based framework that adapts pretrained transformer embeddings to downstream tasks by building a task supervised diffusion geometry over the frozen embeddings, without any backbone training. Applied to antimicrobial peptide design, FDD yields low-dimensional, predictive, and interpretable representations that support property prediction, retrieval, and latent-space interpolation.
Towards Distilling the Representational Geometry of Language Models
Andrew J Steindl
Dhananjay Bhaskar
Ian Adelstein
Large neural networks are increasingly combined through weight-space merging and distilled into smaller models, but weights and activations … (voir plus)are not canonical coordinates: two networks can implement similar functions while differing by permutations, rescalings, or basis changes. We ask whether these operations can instead be guided by the geometry of a model's representation space. For each model, we summarize the relationships among probe examples using a diffusion operator built from internal activations. We then test a simple distillation procedure in which a student is trained to match a fixed diffusion-geometry target derived from a teacher. Across Pythia, BERT, and DeBERTa-v3, this target does not generally make the student inherit the teacher's specific representation geometry. Instead, distilled students usually remain closer to matched task-only controls, and the objective has little effect on held-out performance. These results suggest that representation-geometry objectives must be evaluated against the geometry a student would learn from ordinary training alone.
Gaining Biological Insights through Supervised Data Visualization
Jake S. Rhodes
Marc Girard
Catherine Larochelle
Boaz Lahav
Elsa Brunet-Ratnasingham
Wei Zhang
Adele Cutler
Anhong Zhou
Daniel E. Kaufmann
Alexandre Prat
Kevin R. Moon
Dimensionality reduction-based data visualization is pivotal in comprehending complex biological data. The most common methods, such as PHAT… (voir plus)E, t-SNE, and UMAP, are unsupervised and therefore reflect the dominant structure in the data, which may be independent of expert-provided labels. Here we introduce a supervised data visualization method called RF-PHATE, which integrates expert knowledge for further exploration of the data. RF-PHATE leverages random forests to capture intricate featurelabel relationships. Extracting information from the forest, RF-PHATE generates low-dimensional visualizations that highlight relevant data relationships while disregarding extraneous features. This approach scales to large datasets and applies to classification and regression. We illustrate RF-PHATE’s prowess through three case studies. In a multiple sclerosis study using longitudinal clinical and imaging data, RF-PHATE unveils a sub-group of patients with non-benign relapsingremitting Multiple Sclerosis, demonstrating its aptitude for time-series data. In the context of Raman spectral data, RF-PHATE effectively showcases the impact of antioxidants on diesel exhaust-exposed lung cells, highlighting its proficiency in noisy environments. Furthermore, RF-PHATE aligns established geometric structures with COVID-19 patient outcomes, enriching interpretability in a hierarchical manner. RF-PHATE bridges expert insights and visualizations, promising knowledge generation. Its adaptability, scalability, and noise tolerance underscore its potential for widespread adoption.
Meta-Merging by Checkpoint Nowcasting
Albert Manuel Orozco Camacho
Model merging---the direct combination of parameters from independently fine-tuned networks---offers a way to compose task-specific capabili… (voir plus)ties without retraining or ensemble inference. Existing merge methods are often built from hand-crafted arithmetic or sparsification heuristics, leaving open whether general learned weight-space operators can be repurposed for merging directly. We study this question with NiNo, a pre-trained checkpoint-nowcasting meta-network originally designed to predict near-future training states from short checkpoint histories. We show that pre-trained NiNo can be reused as a data-free pairwise meta-merge operator for independently fine-tuned models. On an 8-task CLIP ViT-B/16 benchmark, NiNo is competitive with strong arithmetic baselines and consistently lands in the same functional region as weight averaging, Task Arithmetic, and TIES. Moreover, NiNo is best on HumanEval in a Qwen3 language extension among the compared merge methods, while extending meta-merge beyond pairs remains an open challenge. These results position learned checkpoint nowcasting as a practical starting point for data-free model merging and motivate future weight-space learners trained for merging explicitly.