Harmonization Shared Autoencoder Gaussian Process Latent Variable Model With Relaxed Hamming Distance.

Area of Science:

• Machine Learning
• Computer Vision
• Data Science

Background:

• Multiview learning outperforms single-view methods in visual classification.
• Gaussian Process Latent Variable Model (GPLVM)-based approaches show strong performance but have limitations.
• Existing methods often fail to harmonize shared latent variables and adapt classifiers effectively, leading to performance degradation.

Purpose of the Study:

• To propose a novel Harmonization Shared Autoencoder GPLVM with Relaxed Hamming Distance (HSAGP-RHD).
• To address limitations in harmonizing shared latent variables and adaptively learning classifiers in multiview learning.
• To enhance the performance of visual classification tasks.

Main Methods:

• An autoencoder structure with Gaussian Process (GP) prior is used to learn shared latent variables.
• A harmonization constraint is embedded to enforce agreement among different views.
• A novel discriminative prior based on Relaxed Hamming Distance (RHD) is introduced for simultaneous feature fusion and classifier learning.

Main Results:

• The proposed HSAGP-RHD method effectively learns shared latent variables and adapts classifiers.
• The RHD-based measurement encourages intra-class compactness and inter-class separability.
• Experimental results on three real-world datasets show superior performance compared to state-of-the-art methods.

Conclusions:

• The HSAGP-RHD model offers a significant advancement in multiview learning for visual classification.
• The novel harmonization constraint and discriminative prior contribute to improved model performance.
• The method demonstrates effectiveness and robustness on diverse datasets.