Sparse alignment for robust tensor learning.

Area of Science:

• Computer Vision
• Pattern Recognition
• Machine Learning
• Multilinear Algebra

Background:

• Multilinear/tensor extensions of manifold learning are prevalent in computer vision and pattern recognition.
• Existing methods often lack robustness and require careful parameter tuning, such as neighborhood size selection.
• A systematic analysis of tensor extensions and alignment techniques is needed to understand their relationship and limitations.

Purpose of the Study:

• To systematically analyze multilinear extensions of manifold learning algorithms using tensor alignment techniques.
• To propose a robust unsupervised tensor learning method for feature extraction.
• To demonstrate the effectiveness of the proposed method compared to existing tensor-based approaches.

Main Methods:

• Developed a general tensor alignment framework by analyzing multilinear extensions of popular manifold learning algorithms.
• Proposed Sparse Tensor Alignment (STA), a novel unsupervised tensor learning method.
• Incorporated L1- and L2-norms into the alignment step of STA to enhance robustness and utilize sparsity for discriminative information.

Main Results:

• The proposed tensor alignment framework reveals intrinsic differences between manifold learning and alignment techniques.
• STA effectively extracts tensor features in an unsupervised manner, demonstrating enhanced robustness.
• Experiments on image, action, and hand gesture databases show STA achieves competitive performance against other unsupervised tensor learning methods.

Conclusions:

• Sparse Tensor Alignment (STA) offers a robust and effective approach for unsupervised tensor feature extraction.
• STA overcomes the challenge of neighborhood size selection inherent in manifold learning-based tensor methods.
• The proposed method shows significant promise for applications in computer vision and pattern recognition.