Alberto Rossi (DIISM, Universities of Florence and Siena)
Oct 4, 2018 – 9:30 AM
DIISM, Artificial Intelligence laboratory (room 201), Siena SI
The ability of develop a complex reasoning on a relational environment is a fundamental characteristic of the human kind, but has been proven to be a difficult task for neural networks.
In this article the author describe how a simple plug-and-play module, called Relational Network (RN), is able to develop a sort of relational reasoning.
They test on three task: visual question answering in which a picture of some geometric object is shown and a (relational or non relational) question about the image is provided.
The second task is a text based question answering based on the bAbI benchmark. At the end they test on dynamic physical systems.
They show how convolutional neural network (or LSTM for text based task) has no chance to solve relational problem. Instead if those well known technique are combined with relational module relational problem can be approached.
Moreover it’s interesting the fact that this module can understand which relation take in account, from the whole set of possible relation, extended the use also to unstructured data.