## Process Flowchart: Neural Architecture Search (NAS) Pipeline ### Overview The image depicts a two-stage technical workflow for a Neural Architecture Search (NAS) system. The process is divided into an "Off-Search Stage" (left) and an iterative "Generate a new architecture" loop (right). The diagram illustrates how hardware configurations and sampled architectures are processed to build a dataset, which then informs an evolutionary search loop to determine the "Best Architecture Selection." ### Components/Axes The diagram is structured into two primary vertical rectangular regions: **1. Left Region: "Off-Search Stage"** * **Input:** "Search Space" * **Process:** "Uniform Sampling" * **Parallel Inputs:** "RPU HW Configurations" and "Dataset of Sampled Architectures" * **Process:** "Training with Noise" * **Output/Storage:** "Dataset Construction" (represented by a database cylinder icon containing four sub-metrics: "Architec.", "Accuracy", "Std", and "AVM"). * **Final Step:** "Training Procedure" (feeds into the right region). **2. Right Region: "Generate a new architecture" (Loop)** * **Start:** "Initial Population Gen." * **State:** "Population" (represented by three neural network icons). * **Evaluation:** "Surrogate Model" and "Evaluation" (parallel processes). * **Decision Point 1:** "AVM < T_AVM" (If "No", loop back to generation; if "Yes", proceed to next step).