Department
Begin
End
Agency
GACR
Identification Code
GA18-21409S
Project Focus
teoretický
Project Type (EU)
other
Publications ÚTIA
Abstract
Anomaly detection, which aims to identity samples very different from majority, is an important tool of unsupervised data analysis. Currently, most methods for anomaly detection use relatively simple shallow models without any complex layers and hierarchies. This in sharp contrast to the area of supervised classification, where hierarchical models with large number of layers stacked on top of each other have proven to be more effective than shallow models. This project aims to partially fill this gap by proposing a systematic study of hierarchical models for anomaly detection from Bayesan and neural network perspectives. Since ve conjecture that the main difference to supervised classification is the need to recognise rare samples, the core of the work focus on how to model frequent and rare samples simultaneously. Models developed during the project will be validated on anomaly detection problems in domains of network security and plasma discharges in Tokamak.