Researchers use semiconductor lasers to achieve high-precision classification

In order to develop photon like brain computing technology, a research team at Strathclyde University integrated spike based neural networks with semiconductor lasers that display the behavior of spike neurons.
They demonstrated high-performance photon spike neural network operations and lower training requirements, and initiated a novel training scheme to improve the results.

This study was published in the scientific collaboration journal 'Intelligent Computing' on August 29, 2023.
Neural networks inspired by biological neural networks have revolutionized artificial intelligence by providing effective solutions for complex tasks. In order to further unleash its potential, scientists have been integrating ideas from other technologies into artificial neural networks and have achieved some success.

A successful example of utilizing the advantages of light is the faster operation speed and higher energy efficiency of optical or photonic neural network systems. This makes them particularly useful for processing large amounts of data and brings hope for future artificial intelligence applications.

The spike neural network proposed by the author relies on a hardware friendly photon system that only includes a vertical cavity surface emitting laser, which is a common device in mobile phones.

This proposal marks a strengthening of previous research conducted by the same group of authors. In their early work, they integrated reservoir computing (an effective technique for building photonic neural networks capable of solving complex tasks) with neural form spike neurons developed using similar lasers.

In the current research, the author completed a highly challenging classification task and adopted alternative training schemes to improve training speed and efficiency while reducing training requirements.

The classification task processed by researchers is a highly complex task, which is a multivariate and nonlinear problem, with each data point containing approximately 500 features and relying on the artificial dataset MADELON.

In order to create a spike neural network, the author utilized an experimental setup that combined the nonlinear spike dynamics of lasers with an architecture inspired by reservoir calculations.

In this architecture, input data is time division multiplexed, i.e. divided into different time slots. Each slot represents a virtual neuron in the neural network.

The input data has been injected into the laser and processed by the laser, and based on whether the input data exceeds a certain threshold, the output is considered as a binary node output, peak or non peak.

The author successfully demonstrated the computational power of photon spike neural networks through traditional least squares regression training techniques and their newly proposed "saliency" training method.
The latter method assigns binary weights to nodes based on their overall utility and importance. Both technologies achieved excellent classification accuracy of over 94%, surpassing established benchmarks in significantly reduced processing time.

Specifically, the accuracy of the new method reached 94.4% and 95.7%, surpassing the results obtained by traditional methods. It is worth noting that even when trained on small datasets containing less than 10 data points, the spike neural network trained using the new method exhibits excellent performance.

The proposed photon spike neural network surpasses traditional digital semiconductor processing systems and has low power consumption, ultra fast performance, and hardware friendly implementation. It can process all virtual nodes using only one laser.

The author believes that this study can create new opportunities for photonics based processing systems that operate entirely on optical hardware, enabling them to handle highly complex tasks in a high-precision, high-speed, and energy-efficient manner.

Source: Laser Network