[Paper Review] Stochastic Learning of Multi-Instance Dictionary for Earth Mover's Distance based Histogram Comparison

Jihong Fan, Ru‐Ze Liang|arXiv (Cornell University)|Sep 3, 2016

Image Retrieval and Classification Techniques1 references3 citations

TL;DR

This paper proposes the first EMD-optimal multi-instance dictionary learning method using stochastic optimization, where a triplet of bags (basic, positive, negative) is used to train a dictionary that minimizes a hinge loss based on Earth Mover's Distance (EMD) comparisons. The method outperforms existing approaches in medical image and natural language retrieval tasks by explicitly optimizing for EMD-based histogram similarity.

ABSTRACT

Dictionary plays an important role in multi-instance data representation. It maps bags of instances to histograms. Earth mover's distance (EMD) is the most effective histogram distance metric for the application of multi-instance retrieval. However, up to now, there is no existing multi-instance dictionary learning methods designed for EMD based histogram comparison. To fill this gap, we develop the first EMD-optimal dictionary learning method using stochastic optimization method. In the stochastic learning framework, we have one triplet of bags, including one basic bag, one positive bag, and one negative bag. These bags are mapped to histograms using a multi-instance dictionary. We argue that the EMD between the basic histogram and the positive histogram should be smaller than that between the basic histogram and the negative histogram. Base on this condition, we design a hinge loss. By minimizing this hinge loss and some regularization terms of the dictionary, we update the dictionary instances. The experiments over multi-instance retrieval applications shows its effectiveness when compared to other dictionary learning methods over the problems of medical image retrieval and natural language relation classification.

Motivation & Objective

• To address the lack of multi-instance dictionary learning methods specifically optimized for Earth Mover’s Distance (EMD) based histogram comparison.
• To overcome the inefficiency of batch optimization in large-scale multi-instance learning by adopting a stochastic learning framework.
• To design a learning objective that ensures the EMD between a basic bag and its positive counterpart is smaller than the EMD to a negative bag, with a margin.
• To improve retrieval performance in multi-instance applications by aligning dictionary learning directly with EMD-based similarity metrics.

Proposed method

• Uses a triplet-based training framework: one basic bag, one positive bag (same class), and one negative bag (different class).
• Maps each bag to a histogram using a multi-instance dictionary, with histogram values derived from instance-dictionary similarities.
• Defines a hinge loss that penalizes cases where the EMD between the basic and negative histogram is not larger than the EMD to the positive histogram by a margin.
• Optimizes the dictionary by minimizing the hinge loss plus an ℓ₂ regularization term on dictionary instances via stochastic gradient descent.
• Leverages the linearity of EMD as a weighted sum of histogram bins to enable efficient gradient computation during optimization.
• Applies stochastic optimization by processing one bag triplet at a time, enabling scalability to large datasets.

Experimental results

Research questions

• RQ1 Can a multi-instance dictionary be learned specifically to optimize for EMD-based histogram comparison, rather than classification error?
• RQ2 Does a stochastic optimization framework improve training efficiency for large-scale multi-instance retrieval?
• RQ3 Can a margin-based hinge loss on EMD distances improve retrieval performance compared to standard dictionary learning methods?
• RQ4 Does EMD-optimized dictionary learning generalize across different domains, such as medical imaging and natural language processing?

Key findings

• The proposed SD-EMD method significantly outperforms MMD, DTD, MILES, and MILIS on the DDSM medical image retrieval dataset in both recall-precision and ROC curves.
• On the SemEval-2010 Task 8 natural language relation classification dataset, SD-EMD again achieved the best performance, demonstrating generalization beyond vision tasks.
• The recall-precision curve of SD-EMD is consistently closer to the top-right corner than all baselines, indicating superior precision and recall trade-offs.
• The ROC curve of SD-EMD is closer to the top-left corner than all other methods, confirming its superior discriminative power in distinguishing positive from negative bags.
• The method’s performance gain is attributed to direct optimization of EMD-based similarity, which is not considered in existing dictionary learning approaches.
• Stochastic optimization enables efficient training on large datasets, avoiding the computational burden of batch processing.