The goal of artificial intelligence is to duplicate the capabilities of the human brain. Deep learning is the branch of AI that gives the machine the power to think like humans. A machine learning expert works with neural networks for a range of applications. Recently, Applied Physics Reviews, a journal by AIP Publishing, published a paper that proposed a new technique for neural network computation. The paper suggests using light instead of electricity to perform the necessary computations. The approach is based on a photonic tensor core that can perform multiplications of matrices in parallel. The methodology is supposed to improve the efficiency and speed of current deep learning paradigms.
To add to the machine learning training, one should be aware of the latest advancements. One such latest entry is photon-based processing units that would enable complex machine learning. In this article, we would look at it comprehensively.
Table of Contents
- Introduction to photon-based processing?
- Working
- Performance
- Final Word
If you are a machine learning novice, you can take up a machine learning course for in-depth knowledge.
Introduction to photon-based processing
Domain-specific coprocessors have emerged as alternatives to centralized paradigms because of the ongoing trend in computing hardware towards increased heterogeneity. The tensor core unit is more reliable than graphic processing units in terms of magnitude, energy efficiency, and signal. Photons have interactive physical properties and local non-volatile mnemonic functionality. A photonic tensor core that can perform tensor operations hasn’t been implemented yet, but several photonic neural network designs have been explored. The new solution strategically utilizes wavelength division multiplexing for photonic parallelism, high two peta-operations-per-second throughputs, and phase changing materials for near-zero static power-consuming novel photonic multi-state memories. This 4-bit photonic tensor core unit’s performance is a result of the combination of physical synergies of material, system, and function. When supported by numerical simulations, the performance can be upgraded 1 order for electrical data. Optically processed data delivers full potential. This shows that photon-based processing units can perform exceptionally well in the future because of the scope of augmenting electronic systems.
Working of Photonic units
Neural networks are trained, such as to perform classification on unseen data. In machine learning, neural networks are trained on unsupervised data and learn to make decisions without any human intervention. After the training, the neural network can find a signature within the data, recognize and classify patterns and objects by producing data inference. The photon-based processing units or photonic TPU processes and stores data in parallel. It features an electro-optical interconnect, allowing the optical memory to be efficiently read and written. It also lets photonic TPU to interface with other architectures. The integrated photonic platforms that integrate efficient optical memory and a tensor processing unit have the same operations but differ in consumption power. Photonic platforms consume only a fraction of the power. It also has higher throughput and can perform inference at the speed of light when opportunely trained.
In an attempt to mimic the human brain, most neural networks stack multiple layers on interconnected neurons. To represent these networks efficiently, a composite function that multiplies vectors and matrices together is used. This representation is required for parallel operations’ performance through vector operation specialized architectures, matrix multiplication, for example. The network’s complexity depends on the intelligence of the task as well as the demanded accuracy of the predictions. More intelligent tasks require more copious amounts of computational data and, in turn, more power for data processing. Currently, graphics processing units (GPUs) or tensor processing units are suitable digital processors for deep learning. Their capabilities are not sufficient when it comes to complex operations and high accuracies. This is because of power limitations and slow transmission of electronic data between the memory and the processor.
Performance
According to an AI developer who authored the published paper, their TPU can outperform an electrical TPU by 2-3 orders. For engines performing intelligent tasks such as 5G with high throughput at the edge of the networks and computing node-distributed networks, photons can be an ideal match. Data signals can exist at the network edges in the form of photons. The photons can be from optical sensors, surveillance cameras, and other sources.
The tensor core unit uses engineered multi-state photonic memories and relies on photonic multiplexed (WDM) signals. The photonic memories can be reprogrammed by changing phases using electrothermal switching. The photonic tensor core can operate as a passive system or can be realized in parallel. There is neither dynamic nor static power dissipation. The runtime complexity is O(1), followed by no additional losses. Additional key features are introduced by avoiding repeaters and optical amplifiers. The execution time of the architecture is limited by the flight of photons in the chip. The advancement in the integration of photonic memories can inherently perform full precision floating-point matrix multiplication and accumulation speeding up intelligent tasks.
Final Word
Photon-based processing units can improve response time, save a tremendous amount of energy, and reduce data center traffic. This means fast processed data for the end-user because most of the data is pre-processed, and only a part needs to be sent to the cloud. An engineered platform that performs fast and energy-efficient matrix multiplication enabling the solving of linear algebraic problems, such as systems of linear equations, inverting matrices, and finding determinants is a need for every Certified Machine Learning Expert and also those enrolled in a machine learning course. Thus, photons can prove to be robust for artificial intelligence.
