A paper in the journal “Applied Physics Reviews” proposes a new method to execute computations required by a neural network, using light instead of using electric power. A photonic tensor core does multiplications of matrices in parallel, enhancing speed and efficiency of the current deep learning paradigms in this approach. Here photons are used to create a more powerful as well as power-efficient processing units for more complex machine learning.
Machine learning done by neural networks is a successful approach for developing artificial intelligence, as researchers aim to replicate brain functionalities for diverse applications. Neural networks were trained to learn for performing unsupervised decision and classification on unobserved data in machine learning. The neural network can produce an inference to recognize and classify objects and patterns and find a signature within the data, once it is trained on data.
The photonic Tensor Processing Units (TPU) stores and processes data in parallel, featuring an electro-optical interconnect, which enables the optical memory to efficiently understand and the photonic Tensor Processing Units to interface with different architectures. Integrated photonic platforms which integrate efficient optical memory can achieve the same operations as a tensor processing unit, but they use a fraction of the power and have higher throughput. It can be used for performing inference at the speed of light when opportunely trained. Most neural networks unravel various layers of interconnected neurons aiming to imitate the human brain. An efficient way to draw these networks is a composite function that multiplies matrices and vectors collectively. This allows the performance of parallel operations through architectures trained in vectorized operations like matrix multiplication.
Current digital processors suitable for deep learning like graphics processing units (GPU) or tensor processing units (TPU) has limits in performing more complicated operations and tasks with more accuracy by the power required to do so and by the creeping transmission of electronic data between the memory and the processor.
The researchers proved that the performance of their new TPU could be 2-3 orders higher than the current electrical TPU. Photons will also be a perfect match for computing node-distributed networks and engines performing further intelligent tasks with high throughput at the edge of networks like 5G.