It is no longer only humans, it’s now machines too which will get smarter by doubting themselves, experiencing failures, and embracing uncertainty. Deep learning, the most powerful approach to artificial intelligence (AI) will have new functionality to make AI systems smarter. The new functionality is known as a sense of uncertainty.
AI researchers at Google and Uber have been introducing changes in two deep learning algorithms to make machines embrace uncertainty. These changes will enable them to know if the machines should second guess themselves while making a prediction or decision. This would be beneficial to improve accuracy in self-driving cars and autonomous machines.
Dustin Tran, one of the researchers working on this approach outlined that if the machine tells how certain it is, it would be essential for safety. On the other hand, if the level of uncertainty is unknown, it would be fatal.
Zoubin Ghahramani, one of the prominent AI researchers opined that uncertainty is vital aspect of human intelligence and rationalization. The addition of this aspect to AI machines would make them smarter and reduce blunders.
Researchers gathered at a major AI conference in Long Beach, California also discussed a new programming language unveiled by Uber, Pyro. This language blended probabilistic programing with deep learning. Noah Goodman, a Stanford University professor affiliated with AI lab of Uber outlined that offering ability to determine probability to deep learning would contribute to making it smarter. A traditional deep learning algorithm learns from data that has been fed, but Pyro has ability to make system preprogrammed with knowledge.
Edward is also a programming language that takes uncertainty into consideration. It has been developed Columbia University and DARPA provided funding for development. Though both languages are in nascent stage of development, tech giants such as Google and Uber attracted attention for them. Both companies are investing enormously in AI research and hiring experts to work on new ideas.
Experts believe that there are a lot of technical challenges in training AI systems to embrace uncertainty and make decisions. But in future, there will be AI systems smarter than they are today.