Graph Based Pattern Recognition
The aim of pattern recognition (PR) is to research methods that are able to (partially) imitate the human capacity of perception and intelligence. PR algorithms are actually employed in various applications (from handwriting recognition to activity predictions for molecular compounds). The question how to represent the underlying data in a formal way is a key issue in PR and related fields. In general, there are two major ways to tackle this step, viz. the statistical approach (using feature vectors) and the structural approach (using graphs). Graphs provide two seminal advantages when compared to feature vectors:
- Graphs are not only able to describe properties, but also relationships among different parts of the underlying data.
- Graphs are not constrained to a fixed size, i.e. the number of nodes and edges can be adapted to the size and the complexity of each individual data object
Due to their power and flexibility, graphs have found widespread application in PR. A prominent example of entities, which can be formally represented in a more suitable and natural way by means of graphs rather than with feature vectors, are chemical compounds.
The field of graph based pattern recognition has a long tradition and can roughly be subdivided into three main eras:
- First era: Graph matching and graph clustering
- Second era: Graph kernels
- Third era: Graph neural networks
This lecture is structured along these three eras (in twelve Chapters). The course includes practical exercises in which the concepts and models presented in the lectures can be implemented and tested on real data.
T3 – Advanced Information Processing |
T6 – Data Science
On successful completion of this class, you will be able to:
Kaspar Riesen |
The course page in ILIAS can be found at https://ilias.unibe.ch/goto_ilias3_unibe_crs_2577195.html.
Schedules and Rooms
|Schedule||Wednesday, 09:15 - 12:00|