Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

nanoML for Human Activity Recognition

Alan T. L. Bacellar, Mugdha P. Jadhao, Shashank Nag, Priscila M. V. Lima, Felipe M. G. Franca, Lizy K. John

TL;DR

This paper tackles the challenge of energy-efficient Human Activity Recognition (HAR) on resource-constrained wearables. It introduces Differentiable Weightless Neural Networks (DWNs), a LUT-based, hardware-friendly nanoML approach trained with Extended Finite Difference and Learnable Mapping, capable of end-to-end optimization. On the UCI-HAR dataset, DWNs achieve 96.34%–96.67% accuracy with per-sample energy as low as 56–104 nJ and end-to-end latency of 5 ns when deployed on an FPGA, while preserving a tiny model footprint. The results demonstrate ultra-low-power edge HAR with substantial energy and memory savings (up to 926,000× energy and 260× memory) compared to CNNs/Transformers, highlighting strong potential for ASIC-enabled wearables and practical edge AI deployment.

Abstract

Human Activity Recognition (HAR) is critical for applications in healthcare, fitness, and IoT, but deploying accurate models on resource-constrained devices remains challenging due to high energy and memory demands. This paper demonstrates the application of Differentiable Weightless Neural Networks (DWNs) to HAR, achieving competitive accuracies of 96.34% and 96.67% while consuming only 56nJ and 104nJ per sample, with an inference time of just 5ns per sample. The DWNs were implemented and evaluated on an FPGA, showcasing their practical feasibility for energy-efficient hardware deployment. DWNs achieve up to 926,000x energy savings and 260x memory reduction compared to state-of-the-art deep learning methods. These results position DWNs as a nano-machine learning nanoML model for HAR, setting a new benchmark in energy efficiency and compactness for edge and wearable devices, paving the way for ultra-efficient edge AI.

nanoML for Human Activity Recognition

TL;DR

This paper tackles the challenge of energy-efficient Human Activity Recognition (HAR) on resource-constrained wearables. It introduces Differentiable Weightless Neural Networks (DWNs), a LUT-based, hardware-friendly nanoML approach trained with Extended Finite Difference and Learnable Mapping, capable of end-to-end optimization. On the UCI-HAR dataset, DWNs achieve 96.34%–96.67% accuracy with per-sample energy as low as 56–104 nJ and end-to-end latency of 5 ns when deployed on an FPGA, while preserving a tiny model footprint. The results demonstrate ultra-low-power edge HAR with substantial energy and memory savings (up to 926,000× energy and 260× memory) compared to CNNs/Transformers, highlighting strong potential for ASIC-enabled wearables and practical edge AI deployment.

Abstract

Human Activity Recognition (HAR) is critical for applications in healthcare, fitness, and IoT, but deploying accurate models on resource-constrained devices remains challenging due to high energy and memory demands. This paper demonstrates the application of Differentiable Weightless Neural Networks (DWNs) to HAR, achieving competitive accuracies of 96.34% and 96.67% while consuming only 56nJ and 104nJ per sample, with an inference time of just 5ns per sample. The DWNs were implemented and evaluated on an FPGA, showcasing their practical feasibility for energy-efficient hardware deployment. DWNs achieve up to 926,000x energy savings and 260x memory reduction compared to state-of-the-art deep learning methods. These results position DWNs as a nano-machine learning nanoML model for HAR, setting a new benchmark in energy efficiency and compactness for edge and wearable devices, paving the way for ultra-efficient edge AI.

Paper Structure

This paper contains 23 sections, 2 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Deep Neural Networks Models for Human Activity Recognition
  3. Preprocessing in Neural Network-Based HAR
  4. State-of-the-art Models
  5. Differentiable Weightless Neural Networks (DWN)
  6. DWN Architecture and Training
  7. Thermometer Unary Encoding
  8. Experimental Evaluation
  9. Dataset Description
  10. Raw Signals
  11. Hand-Crafted Features
  12. Data Splits
  13. Thermometer Encoding
  14. Model Configurations and Training Details
  15. Data Augmentation
  16. ...and 8 more sections

Figures (2)

  • Figure 1: A simple example of a DWN with two layers of LUT-3s. DWNs compute entirely using multiple layers of directly chained lookup tables. Input features are concatenated to form addresses for the first-layer LUTs, whose binary outputs are combined to address the LUTs in subsequent layers. The outputs of the final layer are aggregated to produce activations for each class, without any intermediate arithmetic operations.
  • Figure 2: Confusion matrices for Differentiable Weightless Neural Network (DWN) models evaluated on the UCI-HAR dataset.