As machine learning models get larger and more complex, they require faster and more energy-efficient hardware to perform calculations. Traditional digital computers are struggling to keep up.
An analog optic neural network can perform the same tasks as a digital one, such as image classification or speech recognition, but because the calculations are done using light instead of electrical signals, optic neural networks can run much faster while consuming less energy.
However, these analog devices are prone to hardware errors, which can make calculations less precise. Microscopic defects in hardware components are one cause of these errors. Errors can accumulate rapidly in an optic neural network with many interconnected components.
Due to the basic characteristics of the devices that make up the optic neural network, some error is inevitable even in error correction techniques. A network large enough to be applicable in the real world would be too obscure to be effective.
MIT researchers have overcome this hurdle and found a way to effectively scale an optic neural network. By adding a tiny hardware component to the optical switches that make up the architecture of the network, they can even reduce the uncorrectable errors that would otherwise accumulate in the device.
Their work could provide a super-fast, energy-efficient, analog neural network that can operate with the same accuracy as a digital one. With this technique, as an optical circuit gets larger, the amount of error in its calculations actually decreases.
“This is notable as it goes against the intuition of analog systems, where larger circuits should have higher errors, so that errors set a limit on scalability. This article addresses the scalability problem of these systems at an MIT Electronics Research Lab (RLE) and Quantum Photonics Lab. “Yes,” says lead author Ryan Hamerly, visiting scientist at NTT Research and visiting scientist at NTT Research.
Hamerly’s co-authors are graduate student Saumil Bandyopadhyay and senior author Dirk Englund, associate professor in the MIT Department of Electrical Engineering and Computer Science (EECS), leader of the Quantum Photonics Laboratory, and RLE fellow. Research published Nature Communication.
multiply by light
An optic neural network consists of many interconnected components that function like reprogrammable, adjustable mirrors. These adjustable mirrors are called Mach-Zehnder Inferometers (MZI). Neural network data is encoded into light fired from a laser into the optic neural network.
A typical MZI includes two mirrors and two beam splitters. Light enters from the top of an MZI where it is split into two interfering parts before being recombined by the second beam splitter and then projected from below to the next MZI in the array. Researchers can take advantage of the interference of these optical signals to perform complex linear algebra operations known as matrix multiplication, which is the way neural networks process data.
But errors that can occur with each MZI accumulate quickly as light passes from one device to another. Some bugs can be avoided by pre-defining and setting MZIs so earlier bugs are overridden by later devices in the array.
“It’s a very simple algorithm if you know what the errors are. But these errors are notoriously difficult to detect because you only have access to the inputs and outputs of your chip,” Hamerly says. “This motivated us to see if it was possible to create calibration-free error correction.”
Hamerly and his collaborators have previously demonstrated a mathematical technique that goes a step further. They could successfully spot the errors and adjust the MZIs accordingly, but even that didn’t make the whole error go away.
Due to the basic nature of an MZI, there are situations where it is impossible to set a device so that all light flows from the bottom port to the next MZI. If the device loses some of the light with each step and the array is too large, only very little power will eventually remain.
“There is a fundamental limit to how good a chip can be, even in error correction. MZIs cannot physically perform certain settings that they need to be configured,” he says.
So the team developed a new type of MZI. The researchers added an additional beam splitter to the end of the device, calling it the 3-MZI because it has three beam splitters instead of two. Because of the way this additional beam splitter mixes light, it becomes much easier for an MZI to achieve the setting it needs to send all the light out of the lower port.
More importantly, the additional beam splitter is only a few micrometers in size and is a passive component, so it does not require any extra wiring. Adding additional beam splitters does not significantly change the size of the chip.
Bigger chip, less bugs
When the researchers ran simulations to test their architecture, they found that it could eliminate many of the uncorrectable errors that hinder accuracy. And as the optical neural network grows, the amount of error in the device actually decreases – the opposite of what happens in a device with standard MZIs.
By using 3-MZIs, Hamerly says, they could potentially create a device large enough for commercial use and with a 20x error reduction.
The researchers also developed a variant of the MZI design specifically for associated errors. These are due to manufacturing flaws – if the thickness of a chip is slightly wrong, the MZIs can all be about the same amount, so the errors are pretty much the same. They found a way to change the configuration of an MZI to be resistant to such errors. This technique also increased the bandwidth of the optic neural network so it could run three times faster.
Now that they have demonstrated these techniques using simulations, Hamerly and his collaborators plan to test these approaches on physical hardware and continue to work towards an optic neural network that they can deploy effectively in the real world.
This research is funded in part by a National Science Foundation graduate research fellowship and the U.S. Air Force Office of Scientific Research.