Use of photons to create more powerful and energy efficient processing units for more complex machine learning.
Machine learning performed by neural networks is a popular approach to developing artificial intelligence, as researchers aim to replicate brain function for a variety of applications.
A document in the diary Reviews of applied physics, of AIP Publishing, proposes a new approach to perform the calculations required by a neural network, using light instead of electricity. In this approach, a photonic tensor core performs matrix multiplications in parallel, improving the speed and efficiency of current deep learning paradigms.
In machine learning, neural networks are trained to learn how to make decisions without classification and classification on invisible data. Once a neural network is trained on data, it can produce an inference to recognize and classify objects and models and find a signature within the data.
The photonic TPU stores and processes data in parallel, with an electro-optical interconnection, which allows you to efficiently read and write the optical memory and the photonic TPU to interface with other architectures.
“We discovered that integrated photonic platforms that integrate efficient optical memory can achieve the same operations as a tensor processing unit, but consume a fraction of the power and have a higher throughput and, if properly trained, can be used to perform the inference at the speed of light, “said Mario Miscuglio, one of the authors.
Most neural networks reveal multiple layers of interconnected neurons with the aim of imitating the human brain. An efficient way of representing these networks is a composite function that multiplies matrices and vectors together. This representation allows the execution of parallel operations through architectures specialized in vector operations such as the multiplication of matrices.
However, the smarter the task, the greater it is precision of the desired forecast, the more complex the network becomes. Such networks require more data to calculate and more power to process that data.
Current digital processors suitable for deep learning, such as graphics processing units or tensor processing units, are limited in performing more complex operations with greater precision by the power required to do so and by the slow transmission of electronic data between the processor and the memory.
The researchers showed that their TPU’s performance could be 2-3 orders higher than an electric TPU. Photons can also be an ideal combination for computing networks and motors distributed in nodes that perform intelligent tasks with high throughput at the edge of a network, such as 5G. At the edges of the network, data signals in the form of photons from surveillance cameras, optical sensors and other sources may already exist.
“Specialized photonic processors can save a huge amount of energy, improve response times and reduce data center traffic,” said Miscuglio.
For the end user, this means that data is processed much faster, as a large portion of the data is preprocessed, which means that only a portion of the data needs to be sent to the cloud or data center.
Reference: “Photonic tensor cores for machine learning” by Mario Miscuglio and Volker J. Sorger, 21 July 2020, Reviews of applied physics.
DOI: 10.1063 / 5.0001942