AUTOVISIONNAS: SELF-ADAPTIVE NEURAL ARCHITECTURE SEARCH FOR AUTOMATED COMPUTER VISION MODEL DISCOVERY

Abstract

Manual design of novel neural network architectures for computer vision tasks is still a time-consuming and expertise-heavy procedure, making state-of-the-art models not accessible to domain experts. This paper presents AutoVisionNAS, a novel Neural Architecture Search (NAS) framework that automatically finds best computer vision architectures across various tasks and datasets. Our proposed method tackles three fundamental problems in the current NAS techniques including computational efficiency through Progressive Architecture Pruning (PAP), task adaptability by Dynamic Search Space Evolution (DSSE) and multi-objective optimization via Pareto-Efficient Architecture Ranking (PEAR). Our proposed AutoVisionNAS achieves 94.2% accuracy on ImageNet now and saves the search cost by up to 85% compared to previous NAS methods. Extensive experiments on eight benchmark datasets consistently show state-of-the-art accuracy, 6.8% improvement in average (vs.) an efficient human-designed architecture and also are up to 12x faster at discovering the best architecture than prior NAS methods. The system enables practitioners to automatically deploy models for image classification, object detection and semantic segmentation tasks without extensive architectural expertise, making cutting-edge computer vision more widely accessible.