Invited Talk #2: Weight Space Learning: Learning Representations from Populations of Neural Network Models

This talk introduces weight space learning as an emerging paradigm for learning representations of neural network models themselves by treating the model's weights as a data modality. The central hypothesis is that trained neural network models populate lower-dimensional structures in weight space, which can reveal latent properties of individual models or can be used to synthesize neural network model weights given a suitable prompt.

This talk will provide an overview of recent progress in representation learning over neural network weights, highlight key methodological challenges, including permutation symmetries, scalable architectures, and the transition from small model zoos to learning from populations of neural networks from publicly available model hubs.

Finally, the talk will outline the roadmap toward a foundation model for neural networks, enabling new capabilities for model analysis, generation, adaptation, and control.