AutoNet-Generated Deep Layer-Wise Convex Networks for ECG Classification.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Deep Learning

Background:

• Neural network design is often a time-consuming trial-and-error process.
• Deep neural network loss functions are typically non-convex, complicating optimization.

Purpose of the Study:

• To propose a novel theorem ensuring layer-wise convexity in neural networks.
• To develop an automated algorithm for generating layer-wise convex networks (LCNs).

Main Methods:

• Introduced the Layer-wise Convex Theorem, constraining layers as overdetermined non-linear systems.
• Developed the AutoNet algorithm for end-to-end generation of LCNs.
• Evaluated AutoNet-LCNs on ECG and non-ECG benchmark datasets.

Main Results:

• AutoNet-LCNs achieved superior performance compared to state-of-the-art models on five benchmark datasets.
• Networks were customized for each dataset without manual fine-tuning in under 2 GPU-hours.
• Resulting networks utilized fewer than 5% of the parameters of existing models.

Conclusions:

• The AutoNet-LCN approach significantly reduces model discovery costs.
• This method enables efficient deep learning model training, even in resource-constrained environments.
• The Layer-wise Convex Theorem provides a robust foundation for automated neural network design.