FilterNet: A Many-to-Many Deep Learning Architecture for Time Series Classification

Area of Science:

• Computer Science
• Machine Learning
• Artificial Intelligence

Background:

• Time series classification is crucial for tasks like human activity recognition using sensor data.
• Existing deep learning models, including Convolutional Neural Networks (CNNs) and Long Short-Term Memory (LSTM) networks, have shown promise but require further optimization.
• There is a need for flexible architectures that balance accuracy, efficiency, and segmentation capabilities.

Purpose of the Study:

• To introduce and benchmark FilterNet, a novel deep learning architecture for time series classification.
• To adapt and integrate popular CNN and LSTM motifs into a many-to-many architecture.
• To improve frame-by-frame accuracy, event segmentation, model size, and computational efficiency in activity recognition.

Main Methods:

• Developed FilterNet, a flexible deep learning architecture combining CNN and LSTM components.
• Implemented a many-to-many network structure for enhanced feature processing.
• Evaluated FilterNet variants against existing models on the Opportunity benchmark dataset.
• Investigated the impact of model ensembling and parameter tuning on performance.
• Quantified the precision of discrete event segmentation.

Main Results:

• FilterNet achieved state-of-the-art performance across all measured accuracy and speed metrics on the Opportunity dataset.
• The proposed architecture demonstrated significant improvements in frame-by-frame and event segmentation accuracy.
• FilterNet exhibited enhanced computational efficiency and reduced model size compared to other models.
• Model ensembling and parameter adjustments further optimized performance.
• The models provided high-quality segmentation of discrete events.

Conclusions:

• FilterNet represents a significant advancement in deep learning for time series classification, particularly for activity recognition.
• The architecture's flexibility allows for extensive customization and application to diverse real-world problems.
• FilterNet offers a superior balance of accuracy, segmentation quality, and computational efficiency.
• Further research can explore potential model extensions and broader applications.