Fujitsu envisions AI that thinks outside the box
Tech company Fujitsu aims to develop creative artificial intelligence that works in unfamiliar scenarios, tackling this project with a state-backed institution to try to give Japanese AI research a leg up on competition overseas.
This type of insightful AI in charge of directing crowds at a large event could manage people even when the unexpected occurs, such as a train accident or sudden bad weather. Such an AI handling corporate cybersecurity could estimate the scope of damage when attacked by an unfamiliar computer virus, then take defensive action while keeping systems up and running.
Fujitsu will invest slightly over 2 billion yen ($17.4 million) in the project over five years. The Japanese company will research the application of this next-generation AI while government-backed institution Riken handles the base technology. Researchers and engineers from both parties will join a staff likely to total about 50 people. Fujitsu and Riken hope to reach the product stage in three to five years.
The current AI boom is driven by “deep learning,” in which computers comb troves of past scenarios and compare them with the current one to make the optimal choice. This method lets AIs learn to recognize images, propose medical treatment options, translate languages and play chess and the Japanese game of shogi.
American powerhouses such as IBM and Microsoft have come to use huge numbers of computers to feed AIs massive volumes of learning data. Chinese companies are spending vast sums in the field as well, turning up the pressure. As development takes on the characteristics of a battle in which huge resources are deployed, Japan’s tighter research budgets are leaving the country in the dust.
But AIs taught with deep learning have trouble with unfamiliar scenarios. An image recognizer trained on a huge array of data still struggles to identify an object it has seen only a few times, for example.
Riken’s research team aims to overcome this limit by devising methods to supplement sparse data so the computer can learn appropriately, as well as having the computer use analogous reasoning to predict the effects of its actions in a new scenario. Such methods could let the researchers break free of the heavy resource requirements for AI development.
“Many countries are researching methods for learning based on little data, or for folding in inferences,” said Yutaka Matsuo, a University of Tokyo project associate professor familiar with AI development trends. “The race to implement them is fierce.”