Surrogate Models in Neural Architecture Search

Created by W.Langdon from gp-bibliography.bib Revision:1.9039

@PhdThesis{Kadlecova:thesis,

• author = "Gabriela Kadlecova",
• title = "Surrogate Models in Neural Architecture Search",
• school = "Department of Theoretical Computer Science and Mathematical Logic, Charles University",
• year = "2025",
• address = "Prague",
• keywords = "genetic algorithms, genetic programming, Neural architecture search, ANN, AutoML, surrogate models, zero-cost proxies, evolutionary algorithms, ZCP, GRAF",
• URL = "http://hdl.handle.net/20.500.11956/205192",
• URL = "https://dspace.cuni.cz/handle/20.500.11956/205192?show=full",
• URL = "https://dspace.cuni.cz/bitstream/handle/20.500.11956/205192/140137009.pdf?sequence=3&isAllowed=y",
• URL = "https://dspace.cuni.cz/bitstream/handle/20.500.11956/205192/140137017.pdf",
• size = "74 pages",
• abstract = "This thesis advances neural architecture search (NAS), a subfield of AutoML aiming to discover optimal neural network architectures. We focus on performance predictors; surrogate models that estimate performance without full training, enabling reduction of the search costs of NAS. We first propose info-NAS, a semi-supervised graph neural network surrogate, and extend GNN-based predictors to genetic programming. Next, we uncover a key bias in popular zero-cost proxies and introduce GRAF, a set of interpretable graph features that, combined with zero-cost scores, achieves state- of-the-art correlation with performance. Finally, we present the first surrogates for the expressive grammar-based search space einspace, adapting GRAF-based predictors and introducing language model surrogates. We demonstrate their transferability and effectiveness in reducing search costs. Together, these contributions enhance the efficiency, interpretability, and practical usability of NAS performance predictors.",
• notes = "Supervisor: Roman Neruda",

}

Genetic Programming entries for Gabriela Kadlecova

Citations