Deep representations of sets

Abstract: The field of Deep Learning has experienced tremendous successes in recent years. In Deep Learning, Deep Neural Networks are used to learn from experience and generalize to unseen data by building complex features from raw input data automatically, avoiding the need of handcrafted feature design. However, classical architectures such as Fully-connected Neural Networks are limited in dealing with more complex input structures such as images, sequences, graphs or sets by the required amount of parameters and required training data necessary to train the network. To overcome this limitation and facilitate learning, structural inductive bias can be introduced, encoded explicitly in design-choices of the network. In this dissertation, we focus on Deep Neural Networks dealing with set-structured inputs and propose six different architectures that are suitable to learn deep representations of sets. Each of the networks is built to process input sets of different properties, such as variable-sized input sets, sets containing elements of different modalities, sets containing time series as elements or noisy data containing many missing values. We propose flexible and efficient deep representations of sets for different learning paradigms and evaluate their performance in two fields, where set-structured inputs occur commonly. Firstly, we consider the application of high-level decision making on highways in autonomous driving, where we develop novel Offline Reinforcement Learning algorithms based on set representations. Then, we consider tasks of detection, prediction and clustering in the field personalized medicine, using Supervised Learning and Unsupervised Learning approaches on data ranging from intracranial Electroencephalography to clinical health records. Demonstrating that the proposed networks outperform classical architectures and Machine Learning methods, we bring the field one step closer towards fully-flexible easy-to-use Deep Learning tools that can be used by non-experts without careful and time-consuming engineering

Standort
Deutsche Nationalbibliothek Frankfurt am Main
Umfang
Online-Ressource
Sprache
Englisch
Anmerkungen
Universität Freiburg, Dissertation, 2022

Schlagwort
Neuronales Netz
Maschinelles Lernen
Deep learning
Deep learning
Maschinelles Lernen
Neuronales Netz
Künstliche Intelligenz
Robotik
Medizin

Ereignis
Veröffentlichung
(wo)
Freiburg
(wer)
Universität
(wann)
2022
Urheber

DOI
10.6094/UNIFR/231844
URN
urn:nbn:de:bsz:25-freidok-2318447
Rechteinformation
Open Access; Der Zugriff auf das Objekt ist unbeschränkt möglich.
Letzte Aktualisierung
25.03.2025, 13:55 MEZ

Datenpartner

Dieses Objekt wird bereitgestellt von:
Deutsche Nationalbibliothek. Bei Fragen zum Objekt wenden Sie sich bitte an den Datenpartner.

Entstanden

  • 2022

Ähnliche Objekte (12)