Dirk Englund

Associate Professor of Electrical Engineering and Computer Science at MIT

Photonic Accelerators for Machine Intelligence

Machine learning technology is facing limitations imposed by computational bottlenecks in electronics for tasks involving functions such as vision, games, control, and language processing. In his talk during the NTT Research Summit Upgrade 2020, Dr. Dirk Englund, Associate Professor of Electrical Engineering and Computer Science at MIT, talked about the benefits of using photonic integrated circuits to overcome these bottlenecks.


The goal isn’t to build an optical computer, Dr. Englund said, but rather to use photonic technologies selectively in place of electronics to make advancements in the overall system. “Photonics really is on the rise in computing, but you have to be careful in how you compare it to electronics, and find where the gain is to be had,” he says.


A typical digital artificial neural network consists of an input layer connected to an output layer to create a computable matrix of their values. In traditional machine learning, the values are called upon from memory, which requires a lot of energy. Furthermore, each input value is multiplied by the matrix, resulting in n-squared multiplications. These two methods are unfavorably costly.


Presenting work from 2017 in partnership with NTT Research scientist Dr. Ryan Hamerly, MIT Professor of Physics Marin Soljacic, OPSIS Foundry, and the MIT Quantum Photonics Group, Englund showcases the cost-reducing effects of integrating optics.


“With photonic integrated circuits [PICs], perhaps, we could do that matrix-vector multiplication directly on the PIC itself,“ he says. He highlights two MIT student startups, LightMatter and LightIntelligence, that created optics-based processors capable of processing 64×64 matrices “on the fly” with less than a nanosecond in latency. This was achieved by integrating waveguides, which facilitate optical communication, directly with CMOS electronics.


Dr. Englund lists a few promising theoretical results of using photonic integrated circuits. The first is an overall improvement in the output of machine learning. Because the energy consumption of an optically integrated device scales as n, unlike a digital computer which scales as n squared, the energy consumption overall allows for more gain from the system.


The second pertains to error rates. With equivalent error rates to digital computing, the error rate is limited only by the device model itself, not the imperfections of the device.


Lastly, and possibly the most illustrious of the advancements, is the possibility of reversible and irreversible computing – which would push machine learning technologies past the boundaries of the Landauer limit for digital computing. Currently, CMOS electronics process information irreversibly, meaning bits of data are overwritten (and, effectively, erased) to inform the next bits. Each time this occurs, a small amount of energy is released, in the form of heat. If we were to continue along the path of Moore’s Law with such devices, we would eventually reach energy release levels that are simply not feasible for computing, both in terms of lost energy and intense heat – the Landauer limit. By directing the flow of photons to carry out communication, rather than relying on erasable bits, it is possible to overcome the limit by achieving reversible logic.


Dr. Englund stresses the importance of selectivity when it comes to integrated optics. Alongside Vivienne Sze and Joel Emer of MIT, he created a Digital Optical Neural Network (DONN), that answers the question, “What if we just replaced only the communication part with optics?” He and his colleagues discovered this method is beneficial for larger neural networks that require hefty communication – the kind that would not fit on a single chip. The ability of optics to rapidly travel long distances means the networks could communicate without the restraints of a chip’s physical space. And there are certainly more ways to exploit optical technologies by applying full system simulations to other types of optical neural networks.


For the full transcript of Dirk Englund’s presentation, click here.


Watch Dirk Englund’s full presentation below.

Photonic Accelerators for Machine Intelligence


Dirk Englund,
Associate Professor of Electrical Engineering and Computer Science at MIT