Conditional High-Order Boltzmann Machines for Supervised Relation Learning.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Relation learning is crucial for many vision tasks.
• Unsupervised high-order Boltzmann Machines show potential but lack discriminative power for tasks like face verification.
• Supervised learning with relation class labels is needed for challenging tasks.

Purpose of the Study:

• Introduce supervised relation learning using relation class labels.
• Propose a Conditional High-Order Boltzmann Machine (CHBM) for binary classification of data relations.
• Develop variants (latent CHBM, gated CHBM) for complex relations and feature learning.

Main Methods:

• Developed CHBM by incorporating relation class labels into high-order multiplicative interactions.
• Introduced latent CHBM for joint relation feature learning and classification.
• Introduced gated CHBM to disentangle factors of variation in data relations.
• Utilized approximate factorization for high-order parameter tensors.
• Designed efficient supervised learning algorithms with pretraining and finetuning.

Main Results:

• Proposed CHBM variants effectively learn and classify data relations.
• Models achieved significant performance improvements on invariant recognition, face verification, and action similarity labeling.
• Supervised relation labels were key to enhancing model discriminative ability.

Conclusions:

• Supervised relation learning with CHBM and its variants offers a powerful approach for vision tasks.
• The proposed models demonstrate superior performance compared to unsupervised methods.
• This work advances the field of relation learning in computer vision.