Machine Learning Under Resource Constraints

Fabian Giesecke

Deploying machine learning models in practice requires balancing model quality with constraints on computation, bandwidth, memory, and storage. I will present approaches ranging from parallel algorithms for satellite time series data and learned input selection for bandwidth-limited neural network inference to index-aware decision trees that enable fast search by classification without scanning emtire data catalogs. I will also sketch compact boosted tree ensembles for deployment on devices with extremely limited resources and outline applications across various domains. My talk, hence, connects algorithmic design, data engineering, model development, and applications. The key message is that efficient machine learning in practice requires designing concepts and models together with the systems at hand that store, move, and process the data.

Weitere Infos zur Veranstaltung