Researchers in the field of neural networks have taken the realm of advanced computing another step further. It appears that scientists are now able to unveil the first photonic neural network capable of using light to conduct superfast computing.
Neural networks herald the modern age of computing. Researchers are now using them to create machines capable of a huge range of skills that we thought only humans possess -- object and face recognition, natural language processing, and machine translation, to name a few.
Now the focus of researchers is to push the boundaries of AI and its capabilities. They slowly want to create circuits that operate like neurons. These are called neuromorphic chips. However, the question is, how do we make them faster?
Alexander Tait and his team at Princeton University in New Jersey may have the answer, as they have just created the world's first neuromorphic chip. This allows it to compute at "ultrafast speeds."
This is thanks to optical computing. Photons are seen to have significantly more bandwidth than electrons, meaning they can process data quicker. However, despite these advantages, making these possible will cost a lot of money. This also halted some research in the field like analog signal processing, which demands speed only photonic chips can provide.
Tait's new neural network concept are opening up new opportunities in the field. He said photonic neural networks that utilize silicon photonic platforms can "access new regimes of ultrafast information processing." Potential points of interest include radio, control and even scientific computing.
According to MIT Technology Review, the core challenge is to make an optical device where each node has the same characteristics as a neuron. The nodes will take the form of tiny circular "waveguides" that are carved into a silicon substrate. This allows light to travel around and modulate the output of a laser working at a threshold.
At its core, this means each node in the system is especially sensitive to light. They only work with a specific wavelength, also known as wave division multiplexing. An important question is how this mimics neural behavior, which Tait and his colleagues show equivalency to a device known as a continuous-time recurrent neural network.
He said the results suggest that CTRNNs could be applied to larger silicon photonic neural networks. This holds vast potential in the field because their results say the effective hardware acceleration factor of their neural network is estimated to be 1,960x in the task.
Should research continue, optical computing may be brought to the mainstream populace for the first time. This still depends on how fast the first generation of electronic neuromorphic chips performs.