+********************************************************************** * * * Einladung * * * * Informatik-Oberseminar * * * +********************************************************************** Zeit: Freitag, 22. Dezember 2023, 11.00 Uhr Ort: Gebäude E3, Seminarraum 118, Ahornstr. 55 Referent: Moritz Ibing M.Sc. Lehrstuhl Informatik 8 Thema: Localized Control over the Latent Space of Neural Networks Abstract: Neural networks (NNs) are prevalent today when it comes to analyzing (classifying, segmenting, detecting, etc.) or generating data in all kinds of modalities (text, images, 3D shapes, etc.). They are so useful in these areas, because they have great representation power, while being easy to optimize and generalizing well to unseen data. However, their complexity makes them hard to interpret and modify. Neural networks are usually used to compute a mapping between the data space and a so-called latent space. Often we are interested in local properties of such a mapping. For example, we might want to slightly change the embedding of a data point to achieve a different classification. Such local modifications however are difficult, as NNs usually have globally entangled properties. In this work we will propose ideas how to deal with this problem. Local control is especially of importance for shape representations. It has been shown that NNs are well suited to represent these e.g. as parametric or implicit functions. However, when a global function is used, local supervision is hard to model. We therefore impose additional structure on the latent space of functional representations, making them easier to work with and more expressive. Such a structured representation makes downstream tasks easier, as we are more versatile regarding the shapes we can represent, we can make use of its regularity for the network design, and it allows a compressed encoding that can help to reduce memory consumption. Our focus will be on general shape generation, but we will also present more specific applications like shape completion or super-resolution among others. Our approaches set the state-of-the-art among generative models both in previously used metrics and a newly introduced measure we adapt for this purpose. Es laden ein: die Dozentinnen und Dozenten der Informatik