Communications of the ACM
Brainlike Computers Are a Black Box. Scientists Are Finally Peering Inside
Researchers at the Heinrich Hertz Institute in Germany have found a way to "look" at neural networks in action and see how they draw conclusions.
Credit: Akritasa/Wikimedia Commons
Researchers at the Heinrich Hertz Institute in Germany say they have developed a method for examining neural networks as they operate and visualizing how they reach conclusions.
The Institute's Wojciech Samek and colleagues produced software that goes backward through neural nets to see where a certain decision was reached, and to what degree the decision shaped the results. In this way, researchers can quantify how much individual inputs influence the software's overall conclusion, assigning them a score.
Using this information, scientists can visualize a superimposed mask highlighting the most influential areas.
Samek says this technique could help cut the amount of data required to train neural nets, and probe errors when they occur in results.
Harvard University's Sara Watson stresses ongoing research in this area is essential as the number of decisions algorithms make in people's daily lives continues to grow.
From Science
View Full Article
Abstracts Copyright © 2017 Information Inc., Bethesda, Maryland, USA
No entries found