Matrix Variate Gaussian Mixture Distribution Steered Robust Metric Learning

Lei Luo, Heng Huang

Abstract

Traditional Mahalanobis Metric Learning (MML) often relies on rigid structural assumptions like low-rank or sparsity for the distance matrix $M$. However, real-world data frequently exhibits complex, irregular spatial structures that these priors fail to capture. This paper introduces a novel robust metric learning framework that partitions the distance matrix into blocks and models them using a Matrix Variate Gaussian Mixture Distribution (GMD). By leveraging the universal approximation properties of GMDs, the model learns a data-driven regularizer that adapts to the underlying complexity of the data. Experimental results on face verification and scene recognition benchmarks demonstrate that this flexible approach significantly outperforms state-of-the-art methods.

Introduction: Beyond Rigid Matrix Priors

In machine learning, the effectiveness of similarity measures often hinges on the Mahalanobis distance, defined by a positive semi-definite matrix $M$. To prevent overfitting, researchers typically regularize $M$ using norms that enforce sparsity or low-rankness. While mathematically convenient, these assumptions are often too restrictive for high-dimensional data like facial images under varying conditions.

This work addresses the limitation of simplistic priors by introducing a flexible, data-driven regularizer. Instead of assuming a global structure, the authors propose partitioning the distance matrix and learning its local distribution through a mixture of Gaussians, allowing for a more nuanced representation of the metric space.

Methodology: GMD-Steered Metric Learning

The proposed approach involves three core components:

  1. Matrix Partitioning and Local Modeling: The distance matrix $M$ is divided into $p \times q$ non-overlapping blocks. Each block is treated as a local representation of the matrix's structure.
  2. Matrix Variate GMD Regularization: Each block is modeled by a Gaussian Mixture Distribution. Since GMDs can approximate any continuous density, this regularizer can capture intricate spatial correlations within $M$ that standard norms ignore.
  3. Optimization via EM: The GMD regularizer is integrated into a Large Margin Nearest Neighbor (LMNN) framework. The resulting objective function is optimized using a specialized Expectation-Maximization (EM) algorithm. The E-step estimates component memberships for each block, while the M-step updates the distribution parameters and the distance matrix $M$ via subgradient descent.

Crucially, the authors provide theoretical proof that despite the use of mixture models, the optimization problem remains convex, ensuring convergence to a global optimum.

Experimental Evaluation

The model was rigorously tested on several challenging datasets, including FGNET (aging faces), Labeled Faces in the Wild (LFW), and PubFig.

Key Results

The GMD-steered method consistently outperformed baseline metrics like standard LMNN and its advanced variants. Qualitative results in face retrieval tasks show higher precision in identifying the correct individuals across varying poses and expressions.

Qualitative face retrieval results comparing the proposed method against baselines. Green borders indicate correct matches, while red borders indicate errors.

On the LFW and PubFig datasets, the method demonstrated superior performance in unconstrained face verification. Quantitative analysis via ROC curves shows that the proposed method (labeled "OURMETHOD") achieves higher True Positive Rates across various False Positive Rates compared to established techniques like KISSME, ITML, and LDML.

ROC curves for PubFig and LFW datasets (using both Attribute and SIFT features). The proposed method consistently achieves the highest Area Under the Curve (AUC).

Specifically, on the FGNET database, the method achieved an accuracy of 59.81%, a significant jump from the 47.65% achieved by standard LMNN. It also showed remarkable robustness in cross-sensor tasks, such as matching visible light images to near-infrared face images.

Discussion and Future Outlook

The core innovation of this research is the realization that regularization doesn't need to be synonymous with simplicity. By "steering" the metric learning process with a learned distribution, the model adapts to the actual structure of the data rather than forcing it into a predefined box.

While the method introduces some computational overhead due to the EM iterations, its superior generalization performance justifies the cost for high-stakes applications like face verification. Future work may explore extending this GMD-steered approach to other types of structured learning problems that currently rely on rigid matrix priors.

Summary updated

Abstract

Mahalanobis Metric Learning (MML) has been actively studied recently in machine learning community. Most of existing MML methods aim to learn a powerful Mahalanobis distance for computing similarity of two objects. More recently, multiple methods use matrix norm regularizers to constrain the learned distance matrixMto improve the performance. However, in real applications, the structure of the distance matrix M is complicated and cannot be characterized well by the simple matrix norm. In this paper, we propose a novel robust metric learning method with learning the structure of the distance matrix in a new and natural way. We partition M into blocks and consider each block as a random matrix variate, which is fitted by matrix variate Gaussian mixture distribution. Different from existing methods, our model has no any assumption on M and automatically learns the structure of M from the real data, where the distance matrix M often is neither sparse nor low-rank. We design an effective algorithm to optimize the proposed model and establish the corresponding theoretical guarantee. We conduct extensive evaluations on the real-world data. Experimental results show our method consistently outperforms the related state-of-the-art methods.
BibTeX
@inproceedings{luo2018matrix,
  author = {Lei Luo and Heng Huang},
  title = {{Matrix Variate Gaussian Mixture Distribution Steered Robust Metric Learning}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2018},
  doi = {10.1609/aaai.v32i1.11801},
}
Figure