Keynote Speakers

Day 1 – Dec 10, 2025

Title: Evolving Approaches to Learning on Graphs with (Some) Data at Scale

Speaker: Aditi Krishnapriyan

Abstract: Graph neural networks (GNNs) have demonstrated success in AI applications within the physical sciences, such as in molecular ML tasks like molecular property prediction and creating interatomic potentials for molecular dynamics simulations. GNNs incorporate inductive biases relevant to these tasks, such as locality for molecules, and various methods have been developed to introduce additional biases into GNNs, including symmetry constraints. Recently, advances in large-scale data generation for atomistic modeling have enabled us to conduct systematic studies to evaluate the learning capabilities of different machine learning models, allowing for direct comparisons between GNNs and Transformers. Unlike GNNs, which require a pre-defined graph for each sample, Transformers typically have fewer inductive biases and can operate without needing to define a graph for each new system. We have surprisingly found that many advantageous properties of GNNs can emerge adaptively when using Transformers, offering increased flexibility for modeling diverse molecular structures. I will explore the implications of these findings and discuss their potential impact on the future of graph learning at the intersection of AI and physical sciences. Additionally, I will also more broadly discuss strategies for moving beyond model architecture focused efforts, focusing on how to leverage learned representations with architecture-agnostic techniques for downstream tasks, including knowledge distillation and generative modeling sampling methods inspired by statistical physics.

Bio: Aditi Krishnapriyan is an Assistant Professor of Chemical and Biomolecular Engineering, and of Electrical Engineering and Computer Sciences at UC Berkeley. Her research interests lie in developing methods in machine learning that are driven by the distinct challenges and opportunities in the natural sciences, with particular interest in physics-inspired machine learning methods.

Day 2 – Dec 11, 2025

Geometric Deep Learning for Neural Artifacts: Symmetry-Aware Learning across Trained Model Weights, Internal Representations, and Gradients

Speaker: Haggai Maron

Abstract: The explosive growth of deep learning has created fundamentally new data modalities: the mathematical byproducts of neural network development. Practitioners generate vast quantities of valuable but underutilized data daily—trained model weights, internal representations during inference, and gradients during training. These byproducts, which we term neural artifacts, encode crucial information about model behavior, optimization dynamics, and internal computations. Applying machine learning directly to neural artifacts holds transformative potential: learning on weight spaces could revolutionize model selection, editing, and generation; learning on internal representations could enable efficient hallucination detection and reliability analysis in large language models; learning on gradient spaces could transform optimization and interpretability. Yet existing methods capture only a fraction of this potential because they fail to account for the fundamental symmetries inherent to these data types. We argue that this rich symmetry structure necessitates tools from geometric and equivariant deep learning, and that deploying these tools has the potential to transform how we learn from neural artifacts and significantly impact the deep learning community as a whole. In this talk, I will present the general paradigm and four works advancing it: (1) Equivariant Architectures for Learning in Deep Weight Spaces (ICML 2023), establishing theoretical foundations for symmetry-aware weight space learning; (2) Graph Metanetworks (ICLR 2024), enabling unified, architecture-agnostic processing through computational graph representations; (3) GradMetaNets (NeurIPS 2025), designing symmetry-aware architectures for gradient data; and (4) Neural Message-Passing on Attention Graphs for Hallucination Detection, demonstrating how symmetry-aware methods enhance model reliability and interpretability.

Bio: Haggai Maron is an Assistant Professor and the Robert J. Shillman Fellow at the Faculty of Electrical and Computer Engineering at the Technion and a senior research scientist at NVIDIA Research at NVIDIA’s lab in Tel Aviv. His primary research interest is in machine learning, with a focus on deep learning for structured data.

Day 3 – Dec 12, 2025

Title: Graph Foundation Models at Google

Speaker: Michael Galkin

Abstract: Graph Foundation Models are getting an increasing adoption in 2025 in a variety of applications. In this talk, we’ll focus on both industrial and scientific advancements. In the first part, we’ll provide an overview of the progress on running GFMs in production and introduce Distributed Graph Flow (DGF), an open-source framework for efficient management of planet-scale graph data and how it powers a variety of GFM-related applications at Google. In the second part, we’ll present our novel research results on how graph learning empowers reasoning and structured data understanding capabilities of frontier models.

Bio: Michael Galkin is a Senior Research Scientist at Google Research, specializing in graph learning and reasoning. Before joining Google, he was an AI Research Scientist at Intel Labs in San Diego and a Postdoctoral Fellow at Mila and McGill University.

Day 3 – Dec 12, 2025

Title: Feature Geometry guides Model Design in Deep Learning

Speaker: Melanie Weber

Abstract: The geometry of learned features can provide crucial insights on model design in deep learning. In this talk, we discuss two recent lines of work that reveal how the evolution of learned feature geometry during training both informs and is informed by architecture choices. First, we explore how deep neural networks transform the input data manifold by tracking its evolving geometry through discrete approximations via geometric graphs that encode local similarity structure. Analyzing the graphs’ geometry reveals that as networks train, the models’ nonlinearities drive geometric transformations akin to a discrete Ricci flow. This perspective yields practical insights for early stopping and network depth selection informed by data geometry. The second line of work concerns learning under symmetry, including permutation symmetry in graphs or translation symmetry in images. Group-convolutional architectures can encode such structure as inductive biases, which can enhance model efficiency. However, with increased depth, conventional group convolutions can suffer from instabilities that manifest as loss of feature diversity. A notable example is oversmoothing in graph neural networks. We discuss unitary group convolutions, which provably stabilize feature evolution across layers, enabling the construction of deeper networks that are stable during training.

Bio: Melanie Weber is an Assistant Professor of Applied Mathematics and of Computer Science at Harvard, leading the Geometric Machine Learning Group. Her research focuses on utilizing geometric structure in data for the design of efficient Machine Learning and Optimization methods with provable guarantees.