Prelegent: **Nguyen Hung Son**

2023-11-24 16:15

Distance Metric Learning (DML) is a machine learning approach that aims to

learn a new distance metric from data, which improves the quality of the distance-based

methods including classification, clustering, dimension reduction, kernel-based techniques,

information (e.g. image) retrieval and computer vision. In case of supervised learning prob-

lems, the output of DML are embeddings, where the input data are mapped to improve a

crisp or fuzzy classification process. In other words, DML focuses on searching for a trans-

formation of the original space into a new Euclidean space, in which the distance of objects

belonging to the same class becomes smaller than the distance between objects belonging

to different classes.

This talk summarizes some of the recent developments in research and application of

the distance metric learning method. Each DML algorithm consists of two main parts:

(1) modelling the distance-based constrains in form of an optimisation problem from a given

data set, and

(2) proposing an efficient algorithm for solving the constructed in step (1)

optimisation problem. In this talk, an overview of modelling techniques for classification,

clustering and information retrieval tasks as well as the optimisation techniques including

convex optimisation, different variants of gradient descent (stochastic, projected, ...), Frank-

Wolfe algorithms or Bregman projection will be presented.

The lecture will also cover one of the latest research topics, which is Deep Distance Metric

Learning. The goal of this research is to design a deep neural network that would be able to

tell which objects in different images are visually similar and which are not? Therefore, the

deep distance metric learning approach is actually a DLM method implemented by using

deep neural networks. The advantage of such models is the discovery of highly representative

non-linear embeddings.

Finally, the perspectives of Distance Metric Learning topic and the challenges in the

near future for research and development will be presented.

2023-11-20