Shifting Capsule Networks from the Cloud to the Deep Edge

TL;DR

This paper introduces a quantized implementation of capsule networks optimized for resource-constrained microcontrollers, significantly reducing memory usage while maintaining high accuracy and enabling real-time processing at the edge.

Contribution

It presents an API and framework for running quantized CapsNets on Arm Cortex-M and RISC-V MCUs, extending existing neural network libraries to support capsule operations with minimal accuracy loss.

Findings

75% reduction in memory footprint

CapsNet inference latency under 40 ms on target MCUs

Minimal accuracy loss of 0.07% to 0.18%

Abstract

Capsule networks (CapsNets) are an emerging trend in image processing. In contrast to a convolutional neural network, CapsNets are not vulnerable to object deformation, as the relative spatial information of the objects is preserved across the network. However, their complexity is mainly related to the capsule structure and the dynamic routing mechanism, which makes it almost unreasonable to deploy a CapsNet, in its original form, in a resource-constrained device powered by a small microcontroller (MCU). In an era where intelligence is rapidly shifting from the cloud to the edge, this high complexity imposes serious challenges to the adoption of CapsNets at the very edge. To tackle this issue, we present an API for the execution of quantized CapsNets in Arm Cortex-M and RISC-V MCUs. Our software kernels extend the Arm CMSIS-NN and RISC-V PULP-NN to support capsule operations with 8-bit integers as operands. Along with it, we propose a framework to perform post-training quantization of a CapsNet. Results show a reduction in memory footprint of almost 75%, with accuracy loss ranging from 0.07% to 0.18%. In terms of throughput, our Arm Cortex-M API enables the execution of primary capsule and capsule layers with medium-sized kernels in just 119.94 and 90.60 milliseconds (ms), respectively (STM32H755ZIT6U, Cortex-M7 @ 480 MHz). For the GAP-8 SoC (RISC-V RV32IMCXpulp @ 170 MHz), the latency drops to 7.02 and 38.03 ms, respectively.