TinyCowNet: Memory- and Power-Minimized RNNs Implementable on Tiny Edge Devices for Lifelong Cow Behavior Distribution Estimation

Abstract

Precision livestock farming promises substantial advantages in terms of animal welfare, product quality and reducing methane emissions, but requires continuous and reliable data on the animal's behavior. While systems suitable for use within the barn exist, grazing over long distances poses challenges. Here, we address this issue by proposing an ultra low-power Edge AI device, minimizing data transmission requirements and potentially improving accuracy as compared to classification-based solutions. Namely, we propose cow behavior distribution regression with Recurrent Neural Networks (RNNs), dubbed TinyCowNet, to estimate mixed-label sample spaces. Without quantization, the random search to minimize resources and maximize accuracy shows networks requiring a memory of 76kB on average and offering an accuracy up to 95.7%. These are implementable on a wide range of low-power Micro Controller Units (MCU) and Field Programmable Gate Arrays (FPGA). Furthermore, our proposed post-training full-integer quantization for RNNs combined with power estimation on 45nm CMOS using experimental literature shows a TinyCowNet occupying a memory around $\approx 2$ kB, having a hypothetical power consumption on the order of 200nW, delivering an accuracy of 95.2% and a Matthews correlation coefficient of 0.86. This work paves the way for the future creation of low-cost, highly accurate cow behavior estimation devices with long battery life that reduce the entry barriers currently hindering precision livestock farming outside the barn.

Keywords

• Animal behavior
• application-specific integrated circuits (ASIC)
• field programmable gate arrays (FPGA)
• Internet of Things (IoT)
• Pareto optimization
• recurrent neural networks (RNN)