Genetic Architecture Search for Binarized Neural Networks

In order for deep learning applications to run efficiently on low-power edge devices, including mobile and internet-of-things systems, it is important to reduce their computational and memory requirements. Binarized neural networks have shown promise in this area, but these are typically designed using existing architectures based on floating-point number representations. A more promising approach is to apply network architecture search algorithms to find optimized binarized architectures. In this paper, encoding schemes for the genetic algorithm search of binarized networks are described. The simulation results demonstrate the effectiveness of the proposed method.