Abstract
Selecting a deep learning model is based on evaluating different neural network execution configurations. Database techniques have a lot to contribute to this evaluation. Provenance data adds semantics to the metrics of each configuration, which can help humans in evaluating the models proposed by automatic tools. In large scale, despite the AutoML tools, their execution may incur significant resource costs while exploring the configuration space. Analyzing provenance data during the exploration of configurations, allows for steering configurations during model generation. By analyzing results from initial executions, human intuition may suggest better settings for the hyperparameters in the following ones. The provenance database acts as a supporting data analysis tool complementing visualization frameworks and provides data persistence for model reproducibility. This talk discusses on the role of provenance for model generation in deep learning and presents some experiments of capturing and querying provenance data for benchmark datasets and scientific Physics guided neural networks.
Speaker
Marta Mattoso (COPPE/UFRJ)