MENU 
AI Seminar
Weight Space Learning: Learning Representations of Populations of Neural Networks
Add to Google Calendar
Abstract:
As the number of trained neural network models continues to grow, a fascinating opportunity has emerged to learn from these diverse model populations, known as “model zoos”. In this talk, we will explore the recent advances in “weight spaces learning” aiming to learn representations of model weights and their applications to discriminative and generative downstream tasks. On the discriminative side, such representations enable advanced model analysis such as e.g., the prediction of model performance without requiring access to test data. On the generative side, these representations support the sampling of new, high-performing models for tasks like initialization, transfer learning, and ensemble creation. Attendees will gain insights into not only how weight space learning could be applied in real-world machine learning tasks such as image classification, but also how it could pave a path towards the training of a foundation model of neural networks.
Bio: