Project R-9066

Query languages for Neural Networks (Research)

Neural networks are at the core of many applications of Artificial Intelligence and Data Science. Rather than being set up manually, these networks are automatically trained on very large volumes of training data. Before arriving at an effective network that is suitable for the application at hand, the data scientist will perform many computerized experiments, and try many different variations of data sets. Managing such a process is challenging, taking into account the context of large research projects or enterprises in data analytics, where these processes are happening on a daily basis by several researchers, often using the same or overlapping data sets. What is needed is a data management methodology that allows to describe neural networks and their data sets, as well as to update them, to query them, in short, to manage them effectively. Also the training and execution procedures must be integrated in such a "data science management system". At the core of such a system will be a declarative query language, akin to the situation in relational data management systems. The goal of this project is to design and implement a data model and query language for neural networks, their data sets, and their executions (both forward and backward). If succesful, the proposed research can lead to manageable processes in the development of functional neural networks, as well as to an increased transparancy and explainability of these networks.

01 October 2018 - 02 October 2018