ODMTCNet: An Interpretable Multiview Deep Neural Network Architecture for Feature Representation.

Area of Science:

• Machine Learning
• Computer Vision
• Deep Learning Architectures

Background:

• Deep cascade architectures are widely used but suffer from a "black-box" nature and overfitting issues, especially with limited data.
• Existing models lack interpretability and robust multiview feature representation capabilities.

Purpose of the Study:

• To propose a novel multiview deep neural network (DNN) model, the Optimal Discriminant Multiview Tensor Convolutional Network (ODMTCNet).
• To address the interpretability and overfitting challenges in deep cascade architectures.
• To develop a general platform for effective multiview feature representation.

Main Methods:

• Integration of statistics-guided optimization (SGO) principles with a deep cascade DNN architecture.
• Development of a discriminant multiview tensor convolution strategy.
• Parameter determination for convolutional layers by solving SGO problems, enabling analytical prediction of performance.
• Incorporation of information quality (IQ) for enhanced multiview feature representation.

Main Results:

• ODMTCNet demonstrates superior performance across five diverse datasets (ORL, FERET, ETH-80, Caltech 256, NTU RGB+D 120).
• The model effectively handles various feature types, forming a versatile platform for multiview representation.
• Statistics-guided optimization provides justified knowledge representations and improves model interpretability.

Conclusions:

• ODMTCNet offers a significant advancement over state-of-the-art methods in multiview feature representation.
• The proposed model effectively mitigates the "black-box" problem and overfitting in deep learning.
• The framework's generic nature and effectiveness are validated across multiple datasets and scales.