Evolutionary Approximation and Neural Architecture Search

Approximate computing, Convolutional neural network, Cartesian genetic programming, Neuroevolution, Energy efficiency

Automated neural architecture search (NAS) methods are now employed to routinely deliver high-quality neural network architectures for various challenging data sets and reduce the designers effort. The NAS methods utilizing multi-objective evolutionary algorithms are especially useful when the objective is not only to minimize the network error but also to reduce the number of parameters (weights) or power consumption of the inference phase. We propose a multi-objective NAS method based on Cartesian genetic programming for evolving convolutional neural networks (CNN). The method allows approximate operations to be used in CNNs to reduce the power consumption of a target hardware implementation. During the NAS process, a suitable CNN architecture is evolved together with selecting approximate multipliers to deliver the best trade-offs between accuracy, network size, and power consumption. The most suitable 8 x N-bit approximate multipliers are automatically selected from a library of approximate multipliers. Evolved CNNs are compared with CNNs developed by other NAS methods on the CIFAR-10 and SVHN benchmark problems.