This degree of uncertainty will allow the most powerful Arptifial Intelligence programs to measure their confidence in a prediction or decision, essentially to know when they should doubt themselves.
The deep learning, which is to provide sample data to a neural network, has been a huge success in recent years, allowing the machines to recognize objects in images or transcribe speech almost perfectly. But this one requires a lot of training data and computing power, and it can be surprisingly fragile.
Doubting oneself, however, offers an alternative.
In the field of autonomous driving, for example, where errors can be fatal, the AI should know the degree of certainty of their knowledge. In fact, uncertainty is a key aspect of human reasoning and intelligence. Adding it to AI programs could make them smarter and less likely to make mistakes, says Zoubin Ghahramani, a professor at the University of Cambridge and chief scientist at Uber.
Pyro, for example, is a new programming language launched by Uber that combines deep learning with probabilistic programming. A conventional deep learning system only learns from the data it receives. However, Pyro can also be used to build a preprogrammed knowledge system.
Edward is another programming language that encompasses uncertainty, developed at Columbia University with funds from DARPA.
Both Pyro and Edward are still in the early stages of development, but it’s not hard to see why Uber and Google are interested in them.
Uber uses machine learning in countless areas, from driver routing to pricing and, of course, autonomous cars. The company has invested heavily in AI, hiring several experts working on new ideas. Google bases its entire business on deep learning.
It is still premature to say that this new approach will yield more interesting results than those offered up to now, but it is possible that AI based on uncertainty will end up constituting a small revolution.