DAG SPICER IS expecting a special package soon, but it’s not a Black Friday impulse buy. The fist-sized motor, greened by corrosion, is from a historic room-sized computer intended to ape the human brain. It may also point toward artificial intelligence’s future.
Spicer is senior curator at the Computer History Museum in Mountain View, California. The motor in the mail is from the Mark 1 Perceptron, built by Cornell researcher Frank Rosenblatt in 1958. Rosenblatt’s machine learned to distinguish shapes such as triangles and squares seen through its camera. When shown examples of different shapes, it built “knowledge” using its 512 motors to turn knobs and tune its connections. “It was a major milestone,” says Spicer.
Computers today don’t log their experiences—or ours—using analog parts like the Perceptron’s self-turning knobs. They store and crunch data digitally, using the 1s and 0s of binary numbers. But 11 miles away from the Computer History Museum, a Redwood City, California, startup called Mythic is trying to revive analog computing for artificial intelligence. CEO and cofounder Mike Henry says it’s necessary if we’re to get the full benefits of artificial intelligence in compact devices like phones, cameras, and hearing aids.
Mythic’s analog chips are designed to run artificial neural networks in small devices. MYTHIC
Mythic uses analog chips to run artificial neural networks, or deep-learning software, which drive the recent excitement about AI. The technique requires large volumes of mathematical and memory operations that are taxing for computers—and particularly challenging for small devices with limited chips and battery power. It’s why the most powerful AI systems reside on beefy cloud servers. That’s limiting, because some places AI could be useful have privacy, time, or energy constraints that mean handing off data to a distant computer is impractical.
You might say Mythic’s project is an exercise in time travel. “By the time I went to college analog computers were gone,” says Eli Yablonovitch, a professor at University of California Berkeley who got his first degree in 1967. “This brings back something that had been soundly rejected.” Analog circuits have long been relegated to certain niches, such as radio signal processing.
Henry says internal tests indicate Mythic chips make it possible to run more powerful neural networks in a compact device than a conventional smartphone chip. “This can help deploy deep learning to billions of devices like robots, cars, drones, and phones,” he says.
Google’s AI Wizard Unveils a New Twist on Neural Networks
Henry likes to show the difference his chips could make with a demo in which simulations of his chip and a smartphone chip marketed as tuned for AI run software that spots pedestrians in video from a camera mounted on a car. The chips Mythic has made so far are too small to run a full video processing system. In the demo, Mythic’s chip can spot people from a greater distance, because it doesn’t have to scale down the video to process it. The suggestion is clear: you’ll be more comfortable sharing streets with autonomous vehicles that boast analog inside.
Digital computers work by crunching binary numbers through clockwork-like sequences of arithmetic. Analog computers operate more like a plumbing system, with electrical current in place of water. Electrons flow through a maze of components like amplifiers and resistors that do the work of mathematical operations by changing the current or combining it with others. Measuring the current that emerges from the pipeline reveals the answer.
That approach burns less energy than an equivalent digital device on some tasks because it requires fewer circuits. A Mythic chip can also do all the work of running a neural network without having to tap a device’s memory, which can interfere with other functions. The analog approach isn’t great for everything, not least because it’s more difficult to control noise, which can affect the precision of numbers. But that’s not a problem for running neural networks, which are prized for their ability to make sense of noisy data like images or sound. “Analog math is great for neural networks, but I wouldn’t balance my check book with it,” Henry says.
If analog comes back, it won’t be the first aspect of the Mark 1 Perceptron to get a second life. The machine was one of the earliest examples of a neural network, but the idea was mostly out of favor until the current AI boom started in 2012.
A simulation of Mythic’s chip can identify more objects from a greater distance because it doesn’t have to scale down the video to process it. MYTHIC
Mythic’s analog plumbing is more compact than the Perceptron Mark 1’s motorized knobs. The company’s chips are repurposed flash memory chips like those inside a thumb drive—a hack that turns digital storage into an analog computer.
The hack involves writing out the web of a neural network for a task such as processing video onto the memory chip’s transistors. Data is passed through the network by flowing analog signals around the chip. Those signals are converted back into digital to complete the processing and allow the chip to work inside a conventional digital device. Mythic has a partnership with Fujitsu, which makes flash memory and aims to get customers final chip designs to test next year. The company will initially target the camera market, where applications include consumer gadgets, cars, and surveillance systems.
Mythic hopes its raise-the-dead strategy will keep it alive in a crowded field of companies working on custom silicon for neural networks. Apple and Google have added custom silicon to power neural networks into their latest smartphones.
Yablonovitch of Berkeley guesses that Mythic won’t be the last company that tries to revive analog. He gave a talk this month highlighting the opportune match between analog computing and some of today’s toughest, and most lucrative, computing problems.
“The full potential is even bigger than deep learning,” Yablonovitch says. He says there is evidence analog computers might also help with the notorious traveling-salesman problem, which limits computers planning delivery routes, and in other areas including pharmaceuticals, and investing.
Something that hasn’t changed over the decades since analog computers went out of style is engineers’ fondness for dreaming big. Rosenblatt told the New York Times in 1958 that “perceptrons might be fired to the planets as mechanical space explorers.” Henry has extra-terrestrial hopes, too, saying his chips could help satellites understand what they see. He may be on track to finally prove Rosenblatt right.