VPNET: Variable Projection Networks.

Area of Science:

• Signal Processing
• Machine Learning
• Artificial Intelligence

Background:

• Traditional neural networks can be computationally expensive and lack interpretability.
• Variable Projection (VP) offers a novel approach to designing more efficient and interpretable models.

Purpose of the Study:

• Introduce VPNet, a novel neural network architecture leveraging Variable Projection (VP).
• Evaluate VPNet's performance in signal processing tasks, specifically classification.
• Demonstrate VPNet's advantages in terms of learning speed, accuracy, and computational cost.

Main Methods:

• Developed VPNet, a model-driven neural network architecture incorporating VP operators.
• Applied VPNet to classify a synthetic dataset and real-world electrocardiogram (ECG) signals.
• Compared VPNet against fully connected and 1D convolutional neural networks.

Main Results:

• VPNet achieved fast learning ability and good accuracy on signal classification tasks.
• VPNet demonstrated a low computational cost for both training and inference compared to other architectures.
• The model exhibited learnable features, interpretable parameters, and compact network structures.

Conclusions:

• VPNet presents a promising alternative to existing neural network architectures in signal processing.
• The efficiency and effectiveness of VPNet suggest potential for broad applications in classification, regression, and clustering.
• VPNet's design facilitates interpretable and compact models with reduced computational demands.