Matched Filter Interpretation of CNN Classifiers with Application to HAR

Area of Science:

• Machine Learning
• Deep Learning
• Signal Processing

Background:

• Time series classification is crucial for analyzing sensory data.
• Convolutional Neural Networks (CNNs) are widely used in machine learning.
• Edge inference is gaining importance for low-latency, private data processing.

Purpose of the Study:

• To present a matched filter (MF) interpretation of CNN classifiers.
• To develop a novel MF CNN model for efficient time series classification.
• To enable highly accurate and lightweight models for edge inference.

Main Methods:

• Developed a MF CNN model using Conv1D and GlobalMaxPooling layers.
• Interpreted CNNs as matched filters for automated feature extraction.
• Trained and tested the model on UCI-HAR, WISDM-AR, and MotionSense datasets.

Main Results:

• Achieved 98% average accuracy and 97% F1 score on human activity recognition (HAR).
• Outperformed state-of-the-art HAR methods in accuracy and runtime.
• Model size is under 150 KB with inference times less than 1 ms.

Conclusions:

• The MF interpretation facilitates the development of interpretable, accurate, and computationally efficient CNNs.
• The proposed model is ideal for deployment on mobile devices for various applications.
• This work advances CNN understanding and enables practical edge AI solutions.