AI Machines Can Learn Unsupervised At Speed Of Light, Study Finds
Researchers from the field of artificial intelligence (AI) development have discovered AI machines can learn unsupervised at the speed of light.
Neural networks performed through machine learning has been a popular approach to the development of AI among scientists for some time, with researchers aiming to replicate human brain functionalities for various applications.
Machine learning neural networks is a type of AI that works to replicate functions performed by an organic human brain, with the goal of teaching itself tasks without the need for supervision. A development that has made me feel more than a little sore about my own dismal attempts to learn Spanish.
Now a paper published in the journal Applied Physics Reviews has proposed an innovative new approach to performing computations required by neural networks, with light being used instead of electricity.
From AIP Publishing, this new approach reportedly makes for significant improvements both in terms of the speed and efficiency of neural networks, tackling some of the challenges that have so far hindered current processors from performing more complex operations.
George Washington University researchers found using photons within neural network (tensor) processing units (TPUs) may well overcome such limitations, creating a substantially more powerful and efficient AI.
More advanced tasks require more complex data, meaning the power demand is greater. With this new approach, a photonic tensor core carries out multiplications of matrices in parallel, allowing for processors to conduct more intelligent tasks.
Once neural networks are data trained, they are able to produce an inference to recognise and classify objects and patterns; finding a signature within the data.
The photonic TPU can store and process data in parallel. This also features an electro-optical interconnect, whereby the optical memory can be read and written efficiently, with the photonic TPU capable of interfacing with other architectures.
One of the authors of the paper, Mario Miscuglio, said:
We found that integrated photonic platforms that integrate efficient optical memory can obtain the same operations as a tensor processing unit, but they consume a fraction of the power and have higher throughput and, when opportunely trained, can be used for performing inference at the speed of light.
Miscuglio added that photonic specialised processors have the potential to ‘save a tremendous amount of energy, improve response time and reduce data center traffic’.
The performance of the light TPU was found to be two to three orders higher than a regular electrical TPU, with photons also found to be an ideal match for computing node-distributed networks as well as engines performing advanced tasks with high throughput at the edge of a network, like 5G.
It’s believed data signals could well already exist at network edges, in the form of photons from sources such as CCTV cameras and optical sensors.
If you have a story you want to tell, send it to UNILAD via [email protected]
CreditsApplied Physics Reviews
Applied Physics Reviews