Sign In

Communications of the ACM

ACM TechNews

Radical New Neural Network Design Could Overcome Big Challenges in AI


The trajectories of neural ordinary differential equations.

Researchers at the University of Toronto and the Vector Institute in Canada have redesigned neural networks by replacing the traditional stacked layers of simple computational nodes that work together to find patterns in data with calculus equations.

Credit: David Duvenaud et al.

Researchers at the University of Toronto and the Vector Institute in Canada have redesigned neural networks without traditional stacked layers of simple computational nodes that work together to find patterns in data; the new design replaces the layers with calculus equations.

The researchers, who dubbed this new design an ordinary differential equations (ODE) solver, said it can model continuous change, and changes certain aspects of training for neural networks.

In a traditional neural network, the user has to specify the number of layers at the start of the training, then wait until training is done to find out how accurate the model is. The new method lets the user specify their desired accuracy first, and the network then will find the most efficient way to train itself within that margin of error.

From Technology Review
View Full Article

 

Abstracts Copyright © 2018 Information Inc., Bethesda, Maryland, USA


 

No entries found