Schedule

Subscribe to our calendar for the most recent schedule.

November 21st 1pm CET - Gabriela Kadlecova

Title: GRAF: Performance Prediction with Neural Graph Features

Abstract: Performance prediction is an essential part of the NAS workflow. It helps to accelerate search algorithms, pre-select promising networks, and estimate the latency and energy consumption of an architecture. Recently, zero-cost proxies have emerged as fast and cheap estimators of network performance; yet they often exhibit poor interpretability and inconsistent results over different search spaces and tasks. In this talk, we present GRAF, an interpretable network encoding based on topological features of the neural graph, such as operation counts, path lengths, and node degrees. GRAF is cheap and easy to compute, as it does not require any forward- or backward passes, and matches the correlation of commonly used zero-cost proxies on several different NAS benchmarks. Additionally, GRAF can be utilized alongside zero-cost proxies as input for a tabular performance predictor. We present state-of-the-art results on several NAS benchmarks, hardware tasks, and robustness objectives.

December 5th - Virginia Aglietti

Title: TBA

Abstract: TBA