It goes without saying that no one can predict the future. This includes truly major breakthroughs and the rate at which they will occur are likely to be very unpredictable. This is because there are a lot of different opinions as to which research topics are most important and which approaches are most likely to produce results.
However, the AI experts I’ve spoken to in my new book, Architects of Intelligence, do know more about the current state of the technology, as well as the innovations on the horizon, than virtually anyone else.
As a futurist myself, I predict there will be two types of important AI breakthroughs going forward:
The first involves practical applications of concepts that researchers have already developed.
I think this is where we will see the most immediate progress, and the pace here will be very fast. It is important to note that a lot of this work will be outside the U.S., particularly in China.
China has many advantages when it comes to creating practical AI applications, including access to massive amounts of data as well as a very large number of highly skilled machine learning experts who can take ideas that are developed in the West and quickly turn them into practical applications. So expect lots of progress in areas like facial recognition, fraud detection, self-driving cars, medical imaging, and so forth.
There will also be very important applications to science and medicine. One of the best examples of this is “AlphaFold”, which is DeepMind’s application of its technology to the problem of predicting protein folding. This has very important applications in medical science. One of the researchers I spoke to in Architects of Intelligence, Daphne Koller, has founded a start-up company, insitro, which is focused on using machine learning to discover and develop new drugs. Applications like this could ultimately greatly reduce the cost and time required to bring new drugs to market.
The second category of major breakthroughs will consist of major conceptual advances.
These are the fundamental ideas that will, in turn, underlie future practical advances. The people I spoke to in Architects of Intelligenceare really the top AI researchers in the world, so many of them are focused on these very challenging areas.
One of the most important focuses of current research is unsupervised learning. Nearly all practical AI applications use supervised learning or the process of training algorithms with large sets of labeled data.
For example, a neural network might be trained to recognize photos of animals by giving it millions of images, each labeled with the correct name. This is very powerful when the data is available, but it is obviously not the way people learn. Children somehow are able to learn the names of animals, as well as language and an understanding of the physical world without being given huge amounts of labeled data. They just learn by listening and interacting with their environments in an unsupervised way. Developing techniques for machines to learn in similar ways is one of he most important challenges in AI. Yoshua Bengio is one of the researchers I talked to who has a lot to say about this.
Other huge breakthroughs would be the development of an AI system that has the kind of common sense that humans take for granted. Oren Etzioni of the Allen Institute for AI, for example, heads up a project named Mosiac, which is specifically focused on this challenge. But other researchers who believe strongly in neural network approaches to AI — like Yoshua Bengio, Yann LeCun and Geoff Hinton — believe this is a completely wrong-headed approach! Who is right? No one knows, and that is why the future of AI will be unpredictable and exciting. It is a wide-open field.
Originally published on Quora.