Neural Graphs

Meta-Learning Intelligence in Recurrent Graph Neural Networks

Project Leadership

Project Manager: Yegor Kuznetsov

• Contact: yegor@uw.edu

Description

Recent work has shown recurrent GNNs capable of learning many graph algorithms. However, it’s easy to think of learning itself as a graph algorithm (i.e. backprop operates on computational graphs, and neural networks are structured as graphs) so our project aims to optimize a viable learning algorithm into a GRNN. In essence, we want to directly optimize for Chollet’s definition of intelligence.

Alignment with I2

We’re optimizing a graph for intelligence.