Institutsseminar/2023-07-14 IPD Böhm

Termin in Kalender importieren: iCal (Download)

Vorträge

Vortragende(r)	David Schulmeister
Titel	Hidden Outliers in Manifolds
Vortragstyp	Proposal
Betreuer(in)	Jose Cribeiro
Vortragssprache
Vortragsmodus	in Präsenz
Kurzfassung	Hidden outliers represent instances of disagreement between a full-space and an ensemble. This adversarial nature naturally replicates the subspace behavior that high-dimensional outliers exhibit in reality. Due to this, they have been proven useful for representing complex occurrences like fraud, critical infrastructure failure, and healthcare data, as well as for their use in general outlier detection as the positive class of a self-supervised learner. However, while interesting, hidden outliers' quality highly depends on the number of subspaces selected in the ensemble out of the total possible. Since the number of subspaces increases exponentially with the number of features, this makes high-dimensional applications of Data Analysis, such as Computer Vision, computationally unfeasible. In this thesis, we are going to study the generation of hidden outliers on the embedded data manifold using deep learning techniques to overcome this issue. More precisely, we are going to study the behavior, characteristics, and performance in multiple use-cases of hidden outliers in the data manifold.

Vortragende(r)	Denis Wambold
Titel	Subspace Generative Adversarial Learning for Unsupervised Outlier Detection
Vortragstyp	Proposal
Betreuer(in)	Jose Cribeiro
Vortragssprache
Vortragsmodus	in Präsenz
Kurzfassung	Outlier detection is an important yet challenging task, especially for unlabeled, high-dimensional, datasets. Due to their self-supervised generative nature, Generative Adversarial Networks (GAN) have proven themselves to be one of the most powerful deep learning methods for outlier detection. However, most state-of-the-art GANs for outlier detection share common limitations. Oftentimes we only achieve great results if the model’s hyperparameters are properly tuned or the underlying network structure is adjusted. This optimization is not possible in practice when the data is unlabeled. If not tuned properly, it is not unusual that a state-of-the-art GAN method is outperformed by simpler shallow methods. We propose using a GAN architecture with feature ensemble learning to address hyperparameter sensibility and architectural dependency. This follows the success of feature ensembling in mitigating these problems inside other areas of Deep Learning. This thesis will study the optimization problem, training, and tuning of feature ensemble GANs in an unsupervised scenario, comparing it to other deep generative methods in a similar setting.