Semi-Supervised Adversarial Learning Using LSTM for Human Activity Recognition.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Human Activity Recognition (HAR) models require extensive labeled data.
• Current HAR models exhibit poor performance on anonymous data from new users.
• Acquiring sufficient data for each new user is challenging.

Purpose of the Study:

• To present a semi-supervised adversarial learning method for HAR using Long Short-Term Memory (LSTM).
• To improve HAR model performance by training on both annotated and unannotated data.
• To develop a model adaptable to routine changes and new activities without prior information.

Main Methods:

• Employs semi-supervised adversarial learning with LSTM networks.
• Utilizes both annotated and unannotated data to enhance learning capabilities.
• Incorporates temporal interactive modeling and synchronized LSTM for sequential data prediction.

Main Results:

• Achieved over 98% accuracy on publicly available smart home datasets.
• Demonstrated effective handling of data ambiguity and estimation of heteroscedastic uncertainty.
• Showcased adaptability to changing human activity routines and new activities.

Conclusions:

• The proposed method represents a novel approach to high-level Human Activity Recognition.
• The technique offers a robust solution for HAR challenges, particularly with limited labeled data.
• This method has broad application prospects in smart environments and beyond.