Neural Architecture Search With Randomized Manifold Sampling

Neural network architectures are very critical in determining optimal model performance across various datasets. Traditional approaches to architecture design often involve manual exploration or heuristic methods, which can be time-consuming and resource-intensive. In this article, we propose a randomized neural architecture search (RandNAS) algorithm aimed at efficiently discovering high-performing architectures. Our approach involves embedding the unstructured space of neural architectures within a manifold, enabling constrained exploration of architectures with promising performance potential. By diligently sampling architectures from regions exhibiting high performance values within this constrained space, our algorithm focuses its search efforts on promising candidates. Furthermore, we iteratively calibrate this process using the most promising candidates, enhancing the efficiency of architecture discovery. We demonstrate the effectiveness of our approach by demonstrating competitive performance compared to state-of-the-art methods for both CNN and RNN architectures.