Discriminative Feature Extraction by a Neural Implementation of Canonical Correlation Analysis.

Area of Science:

• Machine Learning
• Computer Vision
• Data Analysis

Background:

• Canonical Correlation Analysis (CCA) extracts features from multiview data but lacks class discriminative ability.
• Existing discriminative CCA (DCCA) methods are often equivalent to single-view LDA and suffer from generalization issues due to sensitivity to outliers.

Purpose of the Study:

• To propose a novel method, discriminative alternating regression (D-AR), for extracting features that are both correlated and class discriminative from multiview data.
• To address the limitations of traditional CCA and DCCA in multiview classification tasks.

Main Methods:

• D-AR employs two alternating multilayer perceptrons with linear hidden layers.
• The perceptrons are trained to predict class labels and each other's outputs, enabling joint learning of correlated and discriminative features.

Main Results:

• Features extracted by D-AR demonstrated significantly higher classification accuracies on test sets.
• Experimental validation was performed on facial expression recognition, object recognition, and image retrieval datasets.

Conclusions:

• D-AR effectively extracts discriminative and correlated features for multiview data.
• The proposed method offers improved performance over existing techniques in classification tasks involving complex datasets.