Artificial Neural Network Ontology (ANNO) / OE Spring 2016

The ANNO Team

Matt Klawonn, MD Ridwan Al Iqbal, Spencer Norris


Welcome! You have found the Artificial Neural Network Ontology, or ANNO. Our goal is to create an ontology to recommend weight initializations for Keras neural network models, with the hope of extending to other artificial neural network libraries.

Our motivation for creating this ontology stems from a lack of theoretical knowledge regarding proper hyperparameter choice in modern Artificial Neural Networks. Many choices for hyperparameter initialization in modern ANNs are best practices: empirical choices informed by some literature. ANNO attempts to capture these best practices along with their provenance to help construct better ANN models.

Our ontology incorporates a number of pre-existing ontologies, including FIBO, SKOS, DCT, and we use Nanopub to capture the provenance of the "best practices."

We expect this ontology to provide a few advantages to machine learning researchers who use neural networks. Firstly, the ontology will be the brain behind a recommendation engine, which users will be able to run over a neural network, pointing out where the network could potentially be improved. Further, the ontology will support its recommendations with provenance, even showing competing recommendations and where they come from should that situation exist.

We highly encourage collaboration and feedback on the ontology, and we will always be looking for researchers to apply the recommendation system to their own neural network models. Please see our getting involved page for more information.

This project is part of the Ontology Engineering class at RPI. The course web page can be found here.

The professor overseeing the work is Prof. Deborah McGuinness.