Representation learning for structured data

školitel: Prof. Ing. RNDr. Martin Holeňa, CSc.
e-mail: zobrazit e-mail
typ práce: dizertační práce
zaměření: MI_MM, MI_AMSM, MINF, APIN
klíčová slova: structured data, neural networks, representation learning, explainability, generative networks
odkaz: http://www2.cs.cas.cz/~martin/
popis: In the last decade, neural learners have become the most popular and most successful kinds of machine learning models, especially their deep variants. After their success in image and nature language processing data, deep learning has spread also to other kinds of data, including structured data, containing combinations of different data types (e.g., numerical and categorical) and/or explicitly indicating semantic relations between different attributes through hierarchies or more general graphs. Yet, learning on structured data still entails many research challenges. Typical supervised and unsupervised learning algorithms are not easily applicable as the structured data are not directly representable in Euclidean space. To this end, representation learning – a specific approach proposed for structured inputs – embeds the data into an Euclidean space, thus yielding a latent representation attempting to preserve the similarity among inputs. Different representations have been already proposed for hierarchical or graph data. However, many open questions in this field still need to be investigated. The topic of the proposed PhD thesis will cover new trends in representation learning on structured data Primarily, it will focus on the following two aspects: 1. Comprehensibility and explainability, which are crucial due to the fact that the black-box nature of typical neural learners hides the semantic content of structured inputs. Representation learning has in this respect a greater potential than traditional methods for the extraction of logical rules from trained neural networks as well as than neuro-symbolic learning, but research into this direction is only emerging. 2. Combing representation learning and generative networks, recently proposed as a means of generating new training and testing data, but not yet sufficiently elaborated for structured inputs. An additional contribution of the thesis will consist in the elaboration of representation learning for the structured data encountered in network security.
literatura: Literature to the topic is extensive, the supervisor will give advice what to read directly to the interested applicant, dependent on background and specific interests.
poznámka: Co-supervisor: Lukáš Bajer, PhD (Cisco Cognitive Intelligence Research and Development)
naposledy změněno: 19.12.2019 10:08:24

za obsah této stránky zodpovídá: Ľubomíra Dvořáková | naposledy změněno: 12.9.2011
Trojanova 13, 120 00 Praha 2, tel. +420 770 127 494
České vysoké učení technické v Praze | Fakulta jaderná a fyzikálně inženýrská | Katedra matematiky