Device and method for training neural network

This describes an automated way to build and train a neural network. It starts by creating many different network designs from a basic template, represented as a flexible string of data. It then uses a first learning step to pick out the most promising designs, which involves exploring various network structures without needing pre-labeled examples. Finally, a second learning step fine-tunes the chosen network using labeled examples to achieve specific performance targets. The claims specify that the network's design and connections are represented by matrices, and the initial design exploration can happen simultaneously for multiple designs.