The researchers trained their experimental device to recognize spoken vowels, a common benchmark task for neural networks. Thanks to the advantages of light, it could do this faster and more efficiently than an electronic device. Other researchers knew that light had the potential to be beneficial for matrix multiplication; the 2017 document showed how to put it into practice.
The study “catalyzed a massive, renewed interest in ONNs,” said Peter McMahon, a photonics expert at Cornell University. “That one was hugely influential.”
Bright ideas
Since that 2017 article, the field has seen steady improvement as various researchers have developed new types of optical computers. Englund and several collaborators recently unveiled a new optical network called HITOP, which combines several advances. More importantly, it aims to increase the computational throughput as a function of time, space and wavelength. Zaijun Chen, a former MIT postdoctoral fellow now based at the University of Southern California, said this helps HITOP overcome one of the drawbacks of optical neural networks: it takes a lot of energy to transfer data from the components electronics to optical components, and vice versa. But by packing information into three dimensions of light, Chen said, it transmits more data faster through the ONN and spreads the energy cost across many calculations. This reduces the cost per calculation. Researchers reported that HITOP could run machine learning models 25,000 times larger than previous chip-based ONNs.
Let’s be clear, the system is still far from living up to its electronic predecessors; HITOP performs about 1 trillion operations per second, while sophisticated Nvidia chips can process 300 times more data, said Chen, who hopes to evolve the technology to make it more competitive. But the efficiency of the optical chip is convincing. “The game here is that we have reduced the cost of energy 1,000 times,” Chen said.
Other groups have created optical computers with different advantages. Last year, a team from the University of Pennsylvania described a new type of ONN that offers unusual flexibility. This chip-based system projects a laser onto a portion of the semiconductor that makes up the electronic chip, which changes the optical properties of the semiconductor. The laser effectively maps the route to be followed by the optical signal and therefore the calculation it performs. This allows researchers to easily reconfigure what the system does. This is a striking difference from most other chip-based optical and electrical systems, whose path is carefully defined in the manufacturing plant and is very difficult to modify.
