After a decade shaped by deep learning, foundation models, and generative AI, future progress in AI does raise fundamental questions about whether current scaling strategies can continue to deliver meaningful advances following the simple recipe: more data, larger models, and more compute. This talk introduces Weight Space Learning, an emerging approach that treats the parameters of trained neural networks as a new data modality and aims to learn neural representations from models, not data. Rather than training from raw images, text, or sensor measurements, we investigate what can be learned directly from populations of neural network models and the knowledge encoded in their weights. This talk will introduce the concepts behind Weight Space Learning, cover potential downstream tasks ranging from model analysis to weight generation, and show that using Weight Space Learning, one could gain a training speedup of up to 20x as compared to conventional neural network training.
Prof. Dr. Damian Borth is Full Professor of Artificial Intelligence and Machine Learning at the University of St. Gallen, where he is part of the School of Computer Science. His research focuses on machine learning, deep learning, artificial intelligence, remote sensing, and Earth observation. Prior to joining St. Gallen, he held research positions at UC Berkeley, the International Computer Science Institute in Berkeley, Columbia University, and the German Research Center for Artificial Intelligence (DFKI). His work has received several distinctions, including the Google Research Scholar Award and the ACM SIGMM Test of Time Award.