Training-Free Proxy-Guided Bayesian NAS for UAV-Constrained TinyML

Neural Architecture Search (NAS) has emerged as a powerful paradigm for automating model design, yet most existing approaches remain training-intensive and computationally prohibitive. In resource-constrained domains such as UAV-based perception and Tiny Machine Learning (TinyML), performing repeated training or fine-tuning during search is infeasible due to strict compute, memory, and energy limitations. We propose a Proxy-Guided Bayesian Optimization NAS framework that eliminates all training during search by modeling a fused set oftrainability proxies(e.g., SynFlow, Jacobian covariance, Neural Tangent Kernel) andhardware proxies(e.g., FLOPs, parameters, latency) within a unified Bayesian surrogate. This surrogate enables uncertainty-aware exploration directly under device-level constraints, guiding the search toward architectures that are both efficient and deployable. Unlike conventional NAS pipelines that demand extensive GPU-time for accuracy evaluations, our method completes the entire search on NATS-Bench (TSS) in only ∼0.8 GPU-hours—achieving a top-1 accuracy of 93.25% with 2.10M parameters, 110M FLOPs, and 0.80 ms latency. This corresponds to an order-of-magnitude reduction in search cost compared to accuracy-driven baselines such as REA and BOHB, while preserving accuracy and satisfying all TinyML deployment budgets (Pmax,Fmax,Lmax). By coupling hardware-awareness with training-free optimization, the proposed approach bridges the gap between proxy-based NAS and real-world, energy-efficient deployment for UAV and edge intelligence applications.