Surprisingly Strong Performance Prediction with Neural Graph Features

Gabriela Kadlecová, Jovita Lukasik, Martin Pilát, Petra Vidnerová, Mahmoud Safari, Roman Neruda, Frank Hutter
Proceedings of the 41st International Conference on Machine Learning, PMLR 235:22771-22816, 2024.

Abstract

Performance prediction has been a key part of the neural architecture search (NAS) process, allowing to speed up NAS algorithms by avoiding resource-consuming network training. Although many performance predictors correlate well with ground truth performance, they require training data in the form of trained networks. Recently, zero-cost proxies have been proposed as an efficient method to estimate network performance without any training. However, they are still poorly understood, exhibit biases with network properties, and their performance is limited. Inspired by the drawbacks of zero-cost proxies, we propose neural graph features (GRAF), simple to compute properties of architectural graphs. GRAF offers fast and interpretable performance prediction while outperforming zero-cost proxies and other common encodings. In combination with other zero-cost proxies, GRAF outperforms most existing performance predictors at a fraction of the cost.

Cite this Paper

BibTeX
@InProceedings{pmlr-v235-kadlecova24a, title = {Surprisingly Strong Performance Prediction with Neural Graph Features}, author = {Kadlecov\'{a}, Gabriela and Lukasik, Jovita and Pil\'{a}t, Martin and Vidnerov\'{a}, Petra and Safari, Mahmoud and Neruda, Roman and Hutter, Frank}, booktitle = {Proceedings of the 41st International Conference on Machine Learning}, pages = {22771--22816}, year = {2024}, editor = {Salakhutdinov, Ruslan and Kolter, Zico and Heller, Katherine and Weller, Adrian and Oliver, Nuria and Scarlett, Jonathan and Berkenkamp, Felix}, volume = {235}, series = {Proceedings of Machine Learning Research}, month = {21--27 Jul}, publisher = {PMLR}, pdf = {https://raw.githubusercontent.com/mlresearch/v235/main/assets/kadlecova24a/kadlecova24a.pdf}, url = {https://proceedings.mlr.press/v235/kadlecova24a.html}, abstract = {Performance prediction has been a key part of the neural architecture search (NAS) process, allowing to speed up NAS algorithms by avoiding resource-consuming network training. Although many performance predictors correlate well with ground truth performance, they require training data in the form of trained networks. Recently, zero-cost proxies have been proposed as an efficient method to estimate network performance without any training. However, they are still poorly understood, exhibit biases with network properties, and their performance is limited. Inspired by the drawbacks of zero-cost proxies, we propose neural graph features (GRAF), simple to compute properties of architectural graphs. GRAF offers fast and interpretable performance prediction while outperforming zero-cost proxies and other common encodings. In combination with other zero-cost proxies, GRAF outperforms most existing performance predictors at a fraction of the cost.} }
Endnote
%0 Conference Paper %T Surprisingly Strong Performance Prediction with Neural Graph Features %A Gabriela Kadlecová %A Jovita Lukasik %A Martin Pilát %A Petra Vidnerová %A Mahmoud Safari %A Roman Neruda %A Frank Hutter %B Proceedings of the 41st International Conference on Machine Learning %C Proceedings of Machine Learning Research %D 2024 %E Ruslan Salakhutdinov %E Zico Kolter %E Katherine Heller %E Adrian Weller %E Nuria Oliver %E Jonathan Scarlett %E Felix Berkenkamp %F pmlr-v235-kadlecova24a %I PMLR %P 22771--22816 %U https://proceedings.mlr.press/v235/kadlecova24a.html %V 235 %X Performance prediction has been a key part of the neural architecture search (NAS) process, allowing to speed up NAS algorithms by avoiding resource-consuming network training. Although many performance predictors correlate well with ground truth performance, they require training data in the form of trained networks. Recently, zero-cost proxies have been proposed as an efficient method to estimate network performance without any training. However, they are still poorly understood, exhibit biases with network properties, and their performance is limited. Inspired by the drawbacks of zero-cost proxies, we propose neural graph features (GRAF), simple to compute properties of architectural graphs. GRAF offers fast and interpretable performance prediction while outperforming zero-cost proxies and other common encodings. In combination with other zero-cost proxies, GRAF outperforms most existing performance predictors at a fraction of the cost.
APA
Kadlecová, G., Lukasik, J., Pilát, M., Vidnerová, P., Safari, M., Neruda, R. & Hutter, F.. (2024). Surprisingly Strong Performance Prediction with Neural Graph Features. Proceedings of the 41st International Conference on Machine Learning, in Proceedings of Machine Learning Research 235:22771-22816 Available from https://proceedings.mlr.press/v235/kadlecova24a.html.

Related Material