Applying MLP-Mixer and gMLP to Human Activity Recognition.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Sensor Data Analysis

Background:

• Deep learning models like Convolutional Neural Networks (CNNs), Transformers, and Multi-layer Perceptrons (MLPs) are increasingly used for human activity recognition (HAR).
• Hyperparameters for sensor data models are often initialized from image processing tasks, potentially leading to oversized models.
• Sensor data typically has lower dimensionality than image data, suggesting potential for model optimization.

Purpose of the Study:

• To investigate the impact of reducing hyperparameters on the performance and efficiency of MLP-based models for HAR.
• To compare the computational efficiency of optimized MLP models against CNNs for sensor-based HAR tasks.

Main Methods:

• Evaluated two MLP-based models: MLP-Mixer and gMLP.
• Reduced hyperparameter values in MLP layers from original recommendations.
• Assessed model performance and computational efficiency on HAR tasks using sensor data.

Main Results:

• MLP-based models' performance showed a positive correlation with the number of parameters.
• Reduced parameter counts in MLP models did not significantly hinder performance for specific HAR tasks.
• MLP-based models demonstrated superior computational efficiency compared to representative CNNs in tested HAR applications.

Conclusions:

• Optimizing hyperparameters for MLP models can lead to more computationally efficient solutions for HAR.
• MLP models offer a promising alternative to CNNs for sensor-based HAR, balancing accuracy with reduced resource requirements.