Mitsubishi Electric develops ‘compact AI’
Mitsubishi Electric has developed what may be a crucial next step in the development of artificial intelligence systems. Its “compact AI” technology eliminates the need for large servers and can be embedded in a far wider scope of devices and machines than existing AI systems can.
The company says that, by filtering information necessary for analysis, the new technology can drastically reduce the processes involved in computation for AI systems.
The development can trim the computation needed for certain tasks by as much as 90%, according to Mitsubishi Electric.
It plans to start offering applications for compact AI technology, such as autonomous driving systems and smarter industrial robots and machine tools, as early as 2017, a company source said.
The breakthrough concerns a branch of AI called deep learning, which attempts to replicate the workings of the layers of neurons in the human brain. Deep-learning software learns to recognize patters in digital representations of sounds, images and other kinds of data.
Until now, one of the biggest hurdles to using artificial neural networks for complicated tasks like driving has been the vast amounts of computation and memory required. This means a huge server is needed to train and operate AI, which often relies on cloud computing.
Such systems consume a significant amount of time in computation and the data communication processes. They are also unsuitable for use where data transmission is difficult, such as in tunnels.
Autonomous driving, for instance, is much more challenging on ordinary roads than on highways, where hazards such as pedestrians and cyclists are not an issue.
“A judgment a person can make in one second takes an onboard artificial intelligence system around 100 seconds,” said Hidetoshi Mishima, an AI expert at Mitsubishi Electric. Onboard AI systems have so far been unable to meet the requirements of autonomous driving in built-up areas.
Compact AI involves an algorithm developed by the company that can automatically select only information necessary for carrying out tasks. The company claims the algorithm can reduce the computational costs and memory requirements for image recognition by 90%.
Resource consumption can be slashed to one-100th of its current level, meaning the system could potentially carry out certain tasks in the same time as would a person.
Artificial neural networks work as systems of interconnected “neurons” that exchange messages between each other. Since neurons transmit signals randomly, the amount of information that needs to be processed is immense. Mitsubishi Electric’s deep-learning algorithm overcomes this problem by filtering out irrelevant information.
By eliminating the need for communications with a large server, an AI system based on this algorithm can be small enough to be installed on board cars or machine tools to carry out related tasks.
Moreoever, not needing to communicate with a remote server makes the system less vulnerable to security risks.
Mitsubishi Electric hopes to roll out products using compact AI as early as 2017. The company has already developed the first application for the new technology — a system that detects careless driving. The system analyzes the driver’s facial information, heartbeat, steering wheel operation and other data to detect signs of low concentration and alert the driver.
The system has proved effective in a test involving 30 drivers, the company said.
“We would also like to apply this technology to satellites, elevators and other products we make,” said Tetsuo Nakakawaji, head of the companay’s Information Technology R&D Center.