Neural architecture search for computer vision tasks

Automated Machine Learning (AutoML) accelerates the deep model design cycle by reducing dependence on manual, expert-crafted archi-tectures. Within this broader field, Neural Architecture Search (NAS) is the subcategory dedicated to automating the generation of neural network architectures. NAS has demonstrated success across various domains, including image classification, natural language processing, and reinforcement learning. Yet, a persistent gap remains between generic NAS benchmarks and the nuanced demands of specialized Computer Vision tasks that require strong generalization under data scarcity, geo-metric consistency, or rapid adaptation to novel scenes. Moreover, the high computational cost of NAS limits its practical adoption, especially in resource-constrained settings. Recent advances in zero-cost and near-zero-cost performance proxies offer promising avenues to alleviate this burden by estimating model quality without extensive training. However, these proxies often struggle to generalize across diverse tasks and complex architectures, highlighting the need for more robust and domain-agnostic metrics. This thesis investigates how to further automate the neural network design process for diverse vision problems through low-cost NAS. The central objective is to evaluate and extend zero-cost or near-zero-cost performance proxies, metrics that can be computed before or after mini-mal training, to guide search over architectural spaces tailored to three challenging domains: (i) deep stereo matching, where structural regular-ization and correspondence reasoning are critical; (ii) generalizable Neural Radiance Fields (NeRF), where view synthesis quality and cross-scene transfer dominate; and (iii) small object detection, where representational efficiency and scale-sensitive feature aggregation are decisive.