Thema der Dissertation:
Interpretable Deep Learning Approaches for the Robust Identification of Peptidoforms in Mass Spectrometry-based Proteomics Thema der Disputation:
Crossmodal learning for zero-shot classification and searching
Interpretable Deep Learning Approaches for the Robust Identification of Peptidoforms in Mass Spectrometry-based Proteomics Thema der Disputation:
Crossmodal learning for zero-shot classification and searching
Abstract: Predictions between two different domains, such as images and text, require flexible deep learning models. To furthermore solve downstream tasks that are not fully pre-specified during training require a framework that learns representations of raw data points, enabling to formulate the specifics of a task after the model has been trained. Radford et al. (2021) pre-train deep models that jointly embed images and text into a joint embedding space. This allows them to perform zero-shot predictions in the notion of classifying images for (during pre-training)- unknown class categories. Not only is their approach competitive with the demonstrated supervised baselines, but it also lays out the groundwork for recent foundation models. In this talk, I will discuss the effectiveness of crossmodal pre-training of deep learning models for task-agnostic settings.
Zeit & Ort
20.03.2025 | 13:00
Seminarraum 005
(Fachbereich Mathematik und Informatik, Takustr. 9, 14195 Berlin)