Thursday, May 4, 2017
The whole world is now excited by the "new AI" based on deep learning. Even Hillary Clinton emphasized it in her new interview. Since I am a grandfather of sorts to this area, I leave soon to give another plenary talk on the subject, at the big IEEE conference in the area (IJCNN2017). Today I turned in the abstract for the following plenary I will give in China (ISNN2017): For many decades, mainstream AI refused to believe that deep learning with neural networks and backpropagation offer true brain like general intelligence, despite numerous successes on tough engineering problems and mathematical advances. The tide changed in 2009 due to a $2 million grant I gave to Ng and LeCun, despite fierce objections which in today's government environment would have prevented the action. Empirical success on well known challenge problems led to follow-ons by DARPA, then by Google, and then a flood of interest by competitors trying to keep up. This past year, new analysis of the best time-series data available for the brain shows that it fits the core principles of backpropagation and deep learning much better than it fits the Hebbian and spiking types of model inherited from the previous century. As the flood of deep learning, internet of things, RNN, BCI, CNN and security technology grows, it may grow out of control. (See www.werbos.com/IT _big_picture.pdf). It is urgent that we move quickly to develop and implement additional new technologies and paradigms, lest the current imbalances and instabilities engulf us all.