Taking AI to the Next Phases
of its Development
The first day’s invited talks were followed by a virtual live panel discussion, moderated by Young Sang Choi, Vice President of Samsung Electronics, and attended by Professor Bengio, Professor LeCun, Professor Finn, Dr. Sainath, Dr. Wortman Vaughan and
Dr. Inyup Kang, President of Samsung Electronics’ System LSI business. “It is my great pleasure to join this Forum,” noted Dr. Kang. “I feel as if I am standing on the shoulders of giants.”
Questions were given to the panel that invited the experts to discuss the ways in which computational bottlenecks can be overcome in order to take AI systems to the next level and be developed to possess the same intelligibility as the human brain. The
panelists weighed the benefits of scaling neural nets as opposed to searching for new algorithms, with Dr. Kang noting that, “We have to try both. Given the scale of human synapses, I doubt that we can achieve the human level of intelligibility
using just current technologies. Eventually we will get there, but we definitely need new algorithms, too.”
Professor LeCun noted how AI research is not just constrained by current scaling methods. “We are missing some major pieces to being able to reach human-level intelligence, or even just animal-level intelligence,” he said, adding that perhaps, in the
near future, we might be able to develop machines that can at least reach the scale of an animal such as a cat. Professor Finn concurred with Professor LeCun. “We still don’t even have the AI capabilities to make a bowl of cereal,” she noted.
“Such basic things are still beyond what our current algorithms are capable of.”
Building on the topic of his invited talk, Professor Bengio added that, in order for future systems to have intelligence comparable to that of the way humans learn as children, a world model will need to be developed that is based on unsupervised learning.
“Our models need to act like human babies in order to go after knowledge in an active way,” he explained.
The panel discussion then moved on to the ways in which the community can bridge the gaps between current technologies and future, human-intelligence level technologies, with all the experts agreeing that there is still much work to be done in developing
systems that mimic the way human synapses work. “A lot of current research directions are trying to address these gaps,” reassured Professor Bengio.
Next, the panel shared their thoughts on how to make AI ‘fairer’ given the inherent biases possessed by today’s societies, with the experts debating the balance that needs to be struck between systems development reform, institutional regulation and corporate
interest. Dr. Wortman Vaughan made the case for introducing a diversity of viewpoints across all parts of the system building process. “I would like to see regulation around processes for people to follow when designing machine learning systems
rather than trying to make everyone meet the same outcomes.”
The final question given to the panel asked for their thoughts on which field will be the next successful application area for end-to-end models. “End-to-end models changed the field of speech recognition by reducing latency and removing the
need for internet connection,” noted Dr. Sainath. “Thanks to this breakthrough, going forward, you’re going to see applications of end-to-end models for such purposes as long meeting transcriptions. We always speak of having ‘one model to
rule them all’, and this is a challenging and interesting research area that has been expanded by the possibilities of end-to-end models as we look to develop a model capable of recognizing all the languages in the world.”