Institutsseminar/2019-04-05

Aus IPD-Institutsseminar
Wechseln zu: Navigation, Suche
Termin (Alle Termine)
Datum Fr 5. April 2019, 11:30 Uhr
Dauer 75 min
Raum Raum 348 (Gebäude 50.34)
Vorheriger Termin Fr 29. März 2019
Nächster Termin Fr 12. April 2019

Vorträge

Vortragende(r) Dennis Vetter
Titel Batch Query Strategies for One-class Active Learning
Vortragstyp Masterarbeit
Betreuer(in) Holger Trittenbach
Kurzfassung Active Learning strategies reduce the amount of labeled observations required by modern machine learning systems. These strategies find queries consisting of the most useful observations and ask a human user for their classification. Querying batches of multiple observations at once can provide various benefits, but also introduces additional complexity to the query selection process. While such batch query strategies exist for Multi-class Active Learning, previous applications of One-class Active learning were focused on sequential query strategies, where only one observation is labeled at a time. This work provide a categorization of existing batch query strategies for multi-class classification. Based on this categorization, we then propose multiple different batch query strategies for One-class Classification. We evaluate them with extensive experiments to determine, to what extend the concepts from Multi-class strategies work with One-class Classification. We then show how these new batch query strategies for One-class classification can be used to drastically reduce learning time without requiring more labeled observations than the use of previous sequential query strategies.
Vortragende(r) Philipp Schüler
Titel Metaheuristics for Query Synthesis in One-Class Active Learning
Vortragstyp Bachelorarbeit
Betreuer(in) Adrian Englhardt
Kurzfassung Active learning describes the topic of a human operator helping with the machine learning process. By asking for a classification of queries, the precision of the machine learning algorithm is increased. Existing research focuses on the idea of using a pool of unlabelled data points or use multiple class cases. We have developed a framework, that allows to synthesize a query in the one-class setting without requiring unlabelled data points. The optimal query is the data point with the highest amount of information. The amount of information for a specific data point is given by the informativeness function. We have created a framework to use metaheuristics to find the maximum of the informativeness function and thus determine the optimal query. We have also conducted experiments to provide a general guideline for the usage of metaheuristics in one-class query synthesis.
Neuen Vortrag erstellen

 

Hinweise