ActiveHARNet: Towards On-Device Deep Bayesian Active Learning for Human Activity Recognition
ActiveHARNet integrates Bayesian neural networks and active learning, significantly reducing labeled data points.
Key Findings
Methodology
ActiveHARNet is a resource-efficient deep ensemble model supporting on-device incremental learning and inference. It represents model uncertainties using dropout approximations in Bayesian Neural Networks, combined with suitable acquisition functions for active learning to select the most informative data points for labeling.
Key Results
- On HHAR and Notch datasets, ActiveHARNet achieves significant efficiency boost during inference, reducing labeled data points by at least 60%.
- Using Variation Ratios acquisition function, user 'i's accuracy improved from 25% to 70%.
- Notch dataset f1-score improved from 0.928 to 0.951, significantly outperforming existing models.
Significance
This research offers a novel solution for human activity recognition in resource-constrained environments, addressing challenges in real-time data labeling and model updating. By reducing labeled data points, it significantly lowers the burden of manual annotation, enhancing model adaptability and efficiency.
Technical Contribution
Introduces a new Bayesian deep learning framework combined with active learning techniques for on-device incremental learning. Compared to existing methods, this model offers significant advantages in resource usage and real-time inference.
Novelty
First to combine Bayesian neural networks with active learning for on-device human activity recognition, providing a new approach for real-time data labeling and model updating.
Limitations
- The model's computational complexity remains high when handling high-dimensional data.
- Performance on certain users is below expectations, requiring further optimization.
Future Work
Future work can explore more acquisition functions and optimization strategies to further improve model efficiency and accuracy.
AI Executive Summary
Human activity recognition is crucial for applications like health monitoring and fall detection. However, existing solutions face challenges in real-time data labeling and model updating. ActiveHARNet integrates Bayesian neural networks and active learning to offer a resource-efficient solution. It represents model uncertainties using dropout approximations and uses acquisition functions to select the most informative data points for labeling. Experimental results show significant efficiency improvements across different users, reducing labeled data points by at least 60%. This research opens new possibilities for human activity recognition in resource-constrained environments, although performance on certain users still requires optimization. Future work will explore more acquisition functions and optimization strategies to further improve model efficiency and accuracy.
Deep Analysis
Background
With the proliferation of mobile and wearable devices, sensor-driven human activity recognition research has evolved. Traditional methods rely on cloud servers and GPUs for computation, but as device computing power increases, on-device machine learning is becoming a new trend.
Core Problem
Real-time data labeling and model updating are core issues in human activity recognition. Existing methods struggle with handling unlabeled data and adapting to new user behaviors.
Innovation
ActiveHARNet combines Bayesian neural networks and active learning techniques to offer a new solution. It represents model uncertainties using dropout approximations and uses acquisition functions to select the most informative data points for labeling.
Methodology
- �� Use Bayesian Neural Networks to represent model uncertainties. • Perform dropout for approximate inference. • Use acquisition functions to select data points for labeling. • Conduct incremental learning and model updating on-device.
Experiments
Experiments conducted on HHAR and Notch datasets using LOOCV strategy to evaluate model performance. Different acquisition functions were used to select labeled data points, assessing model accuracy and efficiency.
Results
ActiveHARNet improved accuracy to 70% on HHAR dataset and f1-score to 0.951 on Notch dataset. Labeled data points reduced by at least 60%.
Applications
The model can be used for real-time health monitoring and fall detection, suitable for resource-constrained wearable devices.
Limitations & Outlook
Performance on certain users is below expectations, requiring further optimization. Computational complexity remains high.
Plain Language Accessible to non-experts
Imagine you're cooking in a kitchen. You need to know which ingredients are most important and which steps require special attention. ActiveHARNet is like a smart chef, able to choose the most important ingredients and adjust the recipe based on real-time conditions with limited resources.
ELI14 Explained like you're 14
Imagine you're playing a game and need to choose the strongest character to win the match. ActiveHARNet is like a smart gamer, able to choose the strongest character with limited resources and adjust strategies based on real-time conditions. It helps you gain an edge in the game!
Glossary
Bayesian Neural Network
A neural network that represents model uncertainties using probability distributions over model parameters.
Used to represent model uncertainties in ActiveHARNet.
Active Learning
A learning method that selects the most informative data points for labeling.
Used for real-time data labeling and model updating.
Dropout
A regularization technique that prevents overfitting by randomly dropping neurons.
Used for approximate inference in Bayesian Neural Networks.
Acquisition Function
A function used to select the most informative data points for labeling.
Used for data point selection in active learning.
Incremental Learning
A learning method that updates the model with real-time data.
Used for on-device model updating.
Open Questions Unanswered questions from this research
- 1 How to further optimize the model to improve performance across all users?
- 2 How to reduce computational complexity to accommodate more devices?
Applications
Immediate Applications
Health Monitoring
Provides more accurate health monitoring through real-time data labeling and model updating.
Long-term Vision
Smart Wearable Devices
Optimizes the model to enhance the performance and user experience of smart wearable devices.
Abstract
Various health-care applications such as assisted living, fall detection etc., require modeling of user behavior through Human Activity Recognition (HAR). HAR using mobile- and wearable-based deep learning algorithms have been on the rise owing to the advancements in pervasive computing. However, there are two other challenges that need to be addressed: first, the deep learning model should support on-device incremental training (model updation) from real-time incoming data points to learn user behavior over time, while also being resource-friendly; second, a suitable ground truthing technique (like Active Learning) should help establish labels on-the-fly while also selecting only the most informative data points to query from an oracle. Hence, in this paper, we propose ActiveHARNet, a resource-efficient deep ensembled model which supports on-device Incremental Learning and inference, with capabilities to represent model uncertainties through approximations in Bayesian Neural Networks using dropout. This is combined with suitable acquisition functions for active learning. Empirical results on two publicly available wrist-worn HAR and fall detection datasets indicate that ActiveHARNet achieves considerable efficiency boost during inference across different users, with a substantially low number of acquired pool points (at least 60% reduction) during incremental learning on both datasets experimented with various acquisition functions, thus demonstrating deployment and Incremental Learning feasibility.