Metric learning for efficient search with high-dimensional multi-modal data

Details

Supervisors

Dupont, Pierre ; Peel, Thomas

Faculty

Ecole polytechnique de Louvain

Degree label

Master [120] : ingénieur civil en informatique

Abstract

Many machine learning algorithms are based on the similarity or distance between objects. For these algorithms, metric learning is a useful preprocessing step to learn a task-specific metric. But until now, metric learning techniques have mainly focused on uni-modal data, while multi-modal data increasingly arise in real-world cases, such as multimedia applications. This master's thesis first explores if taking the multi-modal aspect of the data into account when learning a metric allows to increase the performance of a kNN classifier. For this, multiple ways to combine the modalities are experimented and some problems that can arise with iterative algorithms are tackled, for instance fixing the learning rate. The state-of-the-art metric learning techniques for multi-modal data do not scale well with the size of the training set. This master's thesis explores some approaches to reduce the execution time of these algorithms. It proposes the MKPOE-LR algorithm to tackle the problem of efficiently learning an effective metric for large datasets. It also shows that some other approaches, such as limiting the number of distance constraints, perform well on the classification task at hand while having a significantly smaller execution time.