Computer deep learning systems based on artificial neural network algorithms have become a cutting-edge focus in the field of computer research. The principle is to enable artificial neural network algorithms to learn like the human brain through practice. In addition to being used for facial and voice recognition, it can also search through vast amounts of medical data for diagnostics or explore chemical equations to find potential new drug synthesis methods.
However, the calculations of such systems are very complex and require high machine performance, making it a challenge even for the most powerful existing computers.
However, a research team from MIT and their collaborators have proposed a new method that uses photons instead of electrons for computation. They stated that this approach would significantly improve computational speed and efficiency. Their experimental results were published today in the prestigious journal Nature Photonics, with authors including MIT postdoc Shen Yichen, graduate student Nicholas Harris, Marin Soljačić, Dirk Englund, and eight other collaborators.
Figure: Conceptual diagram of a programmable nano-photonic processor
In fact, many researchers have long promoted photon-based computers. However, Soljačić stated, “They are too optimistic, and the results often backfire.” Although many applications of photon-based computers are impractical, the photon-based neural network system “may be feasible for certain applications of deep learning algorithms,” he said.
Soljačić’s remarks are not without reason. It is well known that artificial intelligence algorithms involve multiple matrix multiplications, which traditional CPU and GPU architectures struggle to handle efficiently. To address this issue, Soljačić and his research team proposed this optical-based computing method after years of research. “The advantage of this optical chip is that once it is set up, the energy consumed for matrix multiplication can theoretically be zero.” Soljačić said, “Although we have not yet developed the entire system, we have already validated the core components.”
Figure: Marin Soljačić
Soljačić likened that even a simple eyeglass lens performs a complex calculation on the light waves passing through it: Fourier transform. The operations executed by the new photon chip, while much simpler than the Fourier transform, are based on a similar principle. This chip uses multi-beam interference technology, and the interference fringe information reflects the results of the required calculations. Researchers refer to this chip as a programmable nano-photonic processor.
Shen Yichen stated that experimental results show that this photon chip architecture can theoretically run artificial intelligence algorithms at faster speeds and with lower power consumption: the energy loss for performing the same operation is even less than one-thousandth of that of traditional chips. “The inherent property of using photons for matrix multiplication is the main reason for the reduction in power consumption and the increase in speed, as dense matrix multiplication is the most time and power-consuming part of artificial intelligence algorithms.”
Figure: Shen Yichen
This nano-photonic processor was developed by Harris and others from Englund’s lab. It is based on a series of interconnected optical waveguides that guide photons, and these connections can be set and programmed as needed to achieve specific computational purposes. “You can program any matrix operation,” Harris said. The team’s ultimate plan is to use multi-layer interleaved devices to implement nonlinear activation function operations, similar to the function of neurons in the brain.
To validate this concept, the team used this programmable nano-photonic processor to implement a neural network algorithm that identifies four basic vowels. The results showed that even using a basic system, they achieved 77% accuracy, while traditional systems achieved 90% accuracy. Soljačić stated that there are “no barriers” to extending the system for higher accuracy.
Englund added that this programmable nano-photonic processor could also have other applications, such as signal processing during data transmission. “High-speed analog signal processing is something this type of processor can accomplish,” and it is faster than methods based on analog-to-digital conversion because light itself is an analog medium. “This method allows operations to be performed directly in the analog domain,” he noted.
Meanwhile, the team also stated that a significant amount of time and effort is still needed to make this system practically applicable. Once this system is successfully scaled and fully operational, it could have many applications, such as in data centers or security systems. This system could also be used in autonomous vehicles or drones. Harris stated, “It can be useful in any situation where large computations are needed but are constrained by time and power limitations.”
