Hardware-Aware Neural Architecture Search

The manual search for optimal neural network topologies, like maximizing the accuracy or minimizing the latency, is a time-consuming process that often requires multiple iterations. Depending on the use case, several target objectives are optimized simultaneously. If the underlying hardware platform needs to be considered as well, the complexity of the search grows even further, with a search space too large for manual search. Neural architecture search automatizes the search for optimal topologies. We will outline the implementation of an evolutionary search algorithm for neural networks, which can take several criteria for the search into consideration. As a final step, hardware awareness is introduced for an exemplary hardware architecture.