End-to-end Evolutionary Neural Architecture Search for Microcontroller Units

Smart wearable devices require accurate, fast, and energy-efficient neural networks to allow for optimal application performance. To advance the field of neural architecture search (NAS), we introduce our end-to-end evolutionary NAS (EvoNAS) for microcontroller units that optimize both, pre-processing and neural network architectures. Each neural network architecture is assessed using the multi-objective accuracy, memory footprint, inference time, and energy consumption, to derive a common performance measure to be maximized. To ensure immediate use of all potential solutions on the microcontroller environment, we create a software-hardware chain in which each neural network is deployed to measure the inference time and power consumption directly. In a proof of concept study, we focused on the analysis of audio-based speech commands. Our experiments suggest that 2D convolutional layers with automatically set pre-processing (short-time Fourier transforms) outperform 1D convolutional layers with raw audio signals. We show that our end-to-end EvoNAS scales with the complexity of the classification task and is still able to find constraint-preserving, and thus deployable, Pareto-optimal neural network architectures even when the classification task is more complex. Our proposed EvoNAS approach is dataset and hardware-agnostic, allowing a universal use across a wide range of applications.