Method for providing a compressed artificial neural network for a combination with hardware for its execution

摘要

The invention relates to a method for optimizing a combination of an artificial neural network and a hardware for its execution. In the method at least one at least pre-trained neural network is provided. Subsequently, a network configuration of at least one neural network is determined in several steps in a plurality of hardware abstraction levels up to reaching a predetermined termination condition, the execution of at least one neural network in the specific network configuration simulated with a virtual hardware model on the respective hardware abstraction level,and in doing so determine a respective estimate of an execution efficiency. According to a predetermined selection criterion, depending on the particular execution efficiencies, at least one of the simulated network configurations is selected and provided after reaching the termination condition for simulation on the next hardware abstraction level or as a final optimized network configuration.