Scientists demonstrate a new optical neural network training method that can crush electronic microprocessors

The current deep neural network system (such as ChatGPT) can quickly improve energy efficiency by 100 times in training, and "future improvements will greatly increase by several orders of magnitude. Scientists from MIT and other institutions have demonstrated a new optical neural network training method that can crush state-of-the-art electronic microprocessors.

Moreover, the computational density of the demonstrated system is about two orders of magnitude higher than that of Nvidia, Google, or Graphcore systems.

Basically, this means that the most advanced models can be trained with 100 times less energy and occupy less space at the same speed.

Artificial neural networks mimic the way biological brains process information. These artificial intelligence systems aim to learn, combine, and summarize information from big datasets, reshaping the field of information processing. Current applications include images, objects, speech recognition, games, medicine, and physical chemistry.

The current artificial intelligence model has reached hundreds of billions of artificial neurons, showing exponential growth and posing challenges to current hardware capabilities.

This paper demonstrates that optical neural network (ONN) methods with high clock speed, parallelism, and low loss data transmission can overcome current limitations.

Our technology opens up a path for large-scale optoelectronic processors to accelerate machine learning tasks from data centers to decentralized edge devices, "the paper wrote.

The ONN method is expected to alleviate the bottlenecks of traditional processors, such as the number of transistors, data mobility energy consumption, and semiconductor size. ONN uses light, which can carry a large amount of information simultaneously due to its wide bandwidth and low data transmission loss. In addition, many photonic circuits can be integrated to expand the system.

In order to move light for calculation, the team led by MIT utilized many laser beams, which were described as "using mass-produced micrometer scale vertical cavity surface emitting lasers for neuron coding".

The researchers explained, "Our scheme is similar to the 'axon synapse dendrite' structure in biological neurons
They believe that the demonstrated system can be expanded through mature wafer level manufacturing processes and photon integration.

Dirk Englund, Associate Professor of Electrical Engineering and Computer Science at the Massachusetts Institute of Technology and the head of this work, explained to SciTechDaily that the size of models such as ChatGPT is limited by the capabilities of today's supercomputers. Therefore, training larger models is not economically feasible.

He claimed, "Our new technology can make it possible to cross machine learning models, otherwise it would not be possible in the near future.

This paper titled "Deep Learning Using Coherent VCSEL Neural Networks" was published by a large team of scientists. This work has received support from the Army Research Office, NTT Research, and NTT Netcast Awards, as well as financial support from the Volkswagen Foundation. The three researchers of the team have applied for patents related to this technology.

Source: Laser Network