Also known as: Neural Architecture Search (NAS) Β· NAS
Neural Architecture Search (NAS) is the automated process of discovering effective neural network architectures for a task. It searches over model structures, such as layer types and connections, often using optimization, reinforcement learning, evolutionary methods, or performance feedback.
A technique for automatically designing the architecture of a neural network. NAS algorithms can reduce the amount of time and resources required to train a neural network. NAS typically uses: - A search space, which is a set of possible architectures. - A fitness function, which is a measure of how well a particular architecture performs on a given task. NAS algorithms often start with a small set of possible architectures and gradually expand the search space as the algorithm learns more about what architectures are effective. The fitness function is typically based on the performance of the architecture on a training set, and the algorithm is typically trained using a reinforcement learning technique. NAS algorithms have proven effective in finding high-performing architectures for a variety of tasks, including image classification, text classification, and machine translation.