Neural Architecture Search (NAS) is an automated approach for designing neural networks by exploring different combinations of layers, connections, operations, and other architectural components. Instead of relying entirely on manual experimentation, a search algorithm evaluates potential network designs and identifies architectures that perform well for a particular task and set of constraints.
For example, NAS can explore different convolutional network structures to find a model that provides strong image-recognition accuracy while using limited computing resources. The search process may consider factors such as accuracy, inference speed, memory consumption, and energy requirements. NAS is used in computer vision, natural language processing, edge AI, and other machine-learning applications where selecting an effective model architecture is important.
