Ultra-compact binary neural networks for human activity recognition on RISC-V processors

Abstract

Human Activity Recognition (HAR) is a relevant inference task in many mobile applications. State-of-the-art HAR at the edge is typically achieved with lightweight machine learning models such as decision trees and Random Forests (RFs), whereas deep learning is less common due to its high computational complexity. In this work, we propose a novel implementation of HAR based on deep neural networks, and precisely on Binary Neural Networks (BNNs), targeting low-power general purpose processors with a RISC-V instruction set. BNNs yield very small memory footprints and low inference complexity, thanks to the replacement of arithmetic operations with bit-wise ones. However, existing BNN implementations on general purpose processors impose constraints tailored to complex computer vision tasks, which result in over-parametrized models for simpler problems like HAR. Therefore, we also introduce a new BNN inference library, which targets ultra-compact models explicitly. With experiments on a single-core RISC-V processor, we show that BNNs trained on two HAR datasets obtain higher classification accuracy compared to a state-of-the-art baseline based on RFs. Furthermore, our BNN reaches the same accuracy of a RF with either less memory (up to 91%) or more energy-efficiency (up to 70%), depending on the complexity of the features extracted by the RF.

Two-stage Human Activity Recognition on Microcontrollers with Decision Trees and CNNs

“…We train our DTs and CNNs in Python, before converting the trained models to optimized C code for the target MCU. For this conversion, we use the library described in [15] for the 1D CNN, while the DT and RF implementations are described in [1]. Both libraries are specifically tailored for the target hardware, leveraging the available SIMD and DSP-oriented instruction set extensions.…”

“…Figure 4 compares our method with a fully tree-based solution, consisting either of a single DT or of a Random Forest (RF) ensemble [1]. Precisely, the green curve shows the Pareto front obtained training all RFs with a number of trees from 1 (single DT) to 15, and depths in the [2,20] interval.…”

“…A HAR task consists of determining the activity performed by an user in a specific time-window of the input signal. This is generally performed through Machine Learning (ML) [1], [2] and more recently Deep Learning (DL) [1], [3], with the latter generally achieving state-of-the-art accuracy.…”

“…On the other hand, DL approaches have shown promising results for HAR [1]- [3], achieving state-of-the-art accuracies (significantly higher than tree-based models) on several datasets. However, most DL models proposed in other works are impossible to deploy on MCUs, requiring > 1 million parameters and floating point operations for inference [3].…”

Channel-wise Mixed-precision Assignment for DNN Inference on Constrained Edge Nodes

“…For instance, it can be noticed that EdMIPS quantizes most of the activations with 8bit, whereas our method exploits its additional flexibility to reduce the activation precision, compensating it with an increase in the bitwidth assigned to an often small subset of the weights channels (e.g., only 3% in c4), in order to obtain the same final accuracy. Eventually, only the first and last layer activations, which notoriously often require higher precision [24] remain at 8bit. Although we report a single example for sake of space, similar considerations apply to the results on the other 3 benchmarks.…”