Samsung recently held an AI Forum event wherein top names in both the company and the field of AI in general spoke to attendees about some of the most mysterious and pressing issues in AI technology today, and more importantly about how they could possibly be solved. Topics ranged from unsupervised learning to optimization and the speakers included MIT Media Lab Professor Cynthia Breazeal, Samsung Research Executive Vice President Sebastian Seung, University of Montreal Professor Yoshua Bengio, and New York University Professor Yann LeCun. The aim of the forum was to open a dialog about the direction of AI going forward and how to overcome some of the biggest barriers to progress that the AI field as a whole currently faces.
LeCun opened up the forum by talking about convolutional neural networks and unsupervised learning, explaining that it tasked AI programs with figuring out for themselves what the best outcome of a given calculation may be. This, in turn, would require many repetitions and massive optimization, but could be used in the future to help AI learn a rudimentary sort of "common sense". The next topic on the list was stochastic gradient descent, an AI optimization convention that's used in the field, but apparently not fully understood just yet. Breaking down its secrets and figuring out why it works as well as it does to optimize AI programs may yield insight into how to improve all optimization techniques, or even the necessary knowledge to develop new ones. Seung, meanwhile, pointed out a study wherein his team used unsupervised learning and a neural network to map out the biological neurons in part of a mouse's brain, a practice that can help lead to neural AI that can change their structure and function to mimic a human brain. Finally, Brezeal closed things out by talking over "social robots" and the ways that advanced robotics and modern AI can be tied together.
Samsung's AI forum provided some measure of insight into where AI technology as a field will be headed in the near future. Many of the conventions, problems and methods discussed in the forum may not come to fruition and actually be applied in consumer-facing technologies for quite some time. This means that these topics, and adjacent ones, will be sitting at the epicenter of an AI arms race of sorts as researchers and companies scramble to solve the problems first, and in the most efficient ways possible.