According to a new study, computers that use light instead of digital switches for calculations can help reduce the energy demand of artificial intelligence (AI). The scientists who invented computers describe it as a new computing paradigm.
Microsoft researchers have developed a prototype analog optical computer (AOC) that can perform several tasks and optimization problems that AI may use.
New computing systems could one day solve certain problems faster, and have less energy than modern digital computers can, the scientists wrote in a study published in Nature on September 3.
“The most important aspect that AOC offers is to estimate an improvement of approximately 100 times more energy efficiency,” said Jannes Gradrow, an AI researcher at Microsoft, in a Microsoft blog post. “That alone is unprecedented in hardware.”
In a new study, Microsoft researchers built a computer that uses microred and camera sensors to perform calculations.
Unlike a regular digital computer that overturns billions of small switches to perform calculations, new systems use different intensities of light and voltage to add and multiply numbers into the feedback loop. The AOC calculates the problem multiple times each last improvement until it reaches a “steady state” or final solution.
Related: The ‘Quantum AI’ algorithm is already surpassing the fastest supercomputers, according to research
You might like it
Because AOC does not convert analog signals into digital signals during calculations, AOC saves energy and overcomes some of the speed limits inherent in digital computing.
This specialized calculation method “will be a special purpose ‘steady-state’ finder for specific AI and optimization problems rather than specific AI and general purpose computers,” Aydogan Ozcan, an optical computing researcher at UCLA who is not involved in the study, told Live Science in an email.
However, for these specific purposes, AOCs may offer significant improvements over digital computing, researchers write in their research.
A new light-based computing paradigm
The team also programmed the “Digital Twin.” This is a computer model that mimics the calculation of physical AOCs. This digital twin can be expanded to handle more variables and more complex calculations.
“The digital twin is where the instrument itself can tackle bigger issues than the instrument itself can tackle right now,” said Michael Hansen, senior director of biomedical signal processing at Microsoft Health Futures, in a blog post.
The team first let AOC perform some simple machine learning tasks, such as image classification. The physical AOC was run in the same way as a digital computer. The team writes that larger AOCs in the future, which can handle more variables, could quickly outperform digital computers in energy efficiency.
Next, researchers reconstructed a 320 x 320 pixel brain scan image using only 62.5% of the original data using AOC digital twins. The digital twin accurately reproduces the scan. This is a feat that scientists could potentially reduce MRI time.
Finally, the team used AOC to solve a set of financial problems that involve minimizing risk while finding the most efficient way to exchange funds between multiple groups.
For now, AOC is a prototype. However, as future models add microred, machines can become much more powerful using millions or billions of variables at once.
“Our goal, our long-term vision is that this is a key part of the future of computing,” says Hitesh Ballani, a researcher with Microsoft’s Cloud Systems Futures team, in a blog post.
Source link
