acm-header
Sign In

Communications of the ACM

ACM TechNews

­Understanding the Limits of Deep Learning


A neural network.

The capabilities of deep-learning neural networks are impressive, but these are tempered by significant constraints.

Credit: nobeastsofierce

The capabilities of deep-learning neural networks are impressive, but these are tempered by significant constraints.

U.S. Defense Advanced Research Projects Agency director John Launchbury sees deep learning occupying part of the second of three waves of artificial intelligence thanks to a "manifold hypothesis," which dictates that different types of high-dimensional natural data tend to cluster and be configured differently when visualized in lower dimensions.

However, Google's Francois Chollet says, "current supervised-perception and reinforcement-learning algorithms require lots of data, are terrible at planning, and are only doing straightforward pattern recognition." Chollet also says scaling up current deep-learning methods will not achieve generalized intelligence. Deep-learning neural networks also can reflect biases and inaccuracies inherent in the data fed to them.

Chollet suggests deep learning's limits could be overcome by initially using "super-human pattern recognition, like deep learning, to augment explicit search and formal systems."

From VentureBeat
View Full Article

 

Abstracts Copyright © 2017 Information Inc., Bethesda, Maryland, USA


 

No entries found