Interaction networks for learning about objects, relations and physics

Adrian Colyer, The Morning Paper, Jan 03, 2017
Commentary by Stephen Downes
files/images/img_0097.jpeg

One of the criticisms of neural networks (and of associative inference generally) is that it cannot generalize. See, for example, Fodor and Pylyshyn 1988. Of course in the 25 years since the criticism was leveled they have faced the sternest of all critics: empirical evidence to the contrary. This paper describes a neural net that can learn Newtonian physics. "Our results provide surprisingly strong evidence of IN’s ability to learn accurate physical simulations and generalize their training to novel systems with different numbers and configurations of objects and relations."

Views: 1 today, 574 total (since January 1, 2017).[Direct Link]
Creative Commons License. gRSShopper

Copyright 2015 Stephen Downes ~ Contact: stephen@downes.ca
This page generated by gRSShopper.
Last Updated: Oct 24, 2017 12:07 a.m.