A new data-driven approach for the optimization of virtual assistants for neural networks

Abstract

Vann (Virual Assistant for Neural Networks) is a software that I developed in collaboration with the DLR research centre (Deutsches Zentrum f¨ur Luft und Raumfahr), aiming at simplifying and enhancing the design process of neural network architectures without imposing any sort of constraint on the model definitions and the data workflow. The outcomes of this system is related to the instantiation of single or multiple neural networks which ultimately solve the user’s problem, by obtaining better target values than those achievable manually. The software is designed as a virtual assistant from which the user can retrieve different types of information on important model design features and make decisions based on a completely data-driven approach. While it is true that the data science industry as a whole cannot be fully automated nowadays, there is still room to automate several processes of every big data-based statistical problem. Through VANN, I tried to verify if a semi-automated system could be an effective response to address and minimize the cognitive overload on experimenters caused by redundant operations and trial-and-error iteration.