Watch a 39-minute AutoML Seminar presentation exploring GRAF, a novel approach to neural network performance prediction using topological graph features. Learn how this interpretable network encoding method leverages operation counts, path lengths, and node degrees to estimate architecture performance without requiring forward or backward passes. Discover how GRAF matches correlation levels of existing zero-cost proxies across multiple NAS benchmarks while offering improved interpretability and consistency. Follow along as speaker Gabi Kadlecová demonstrates state-of-the-art results when applying GRAF to various NAS benchmarks, hardware tasks, and robustness objectives, showing how it can be effectively combined with zero-cost proxies for enhanced tabular performance prediction.