Neural networks emulate the distributed structure of a biological brain to achieve cognitive abilities comparable to humans, but in significantly less time. These technologies are now the foundation of machine learning and AI systems that can analyze and adjust to their surroundings, operating autonomously. Their applications span a variety of fields, including image and speech recognition and synthesis, bioinformatics, molecular sequencing, high-performance computing, and more.
Unlike traditional computing methods, neural networks require significant initial training with a vast amount of preexisting knowledge, which they then adapt through experience. However, training is an incredibly energy-intensive process, and with the exponential growth in computing power, neural networks’ energy consumption is also rapidly increasing, doubling every six months.
To address this issue, photonic circuits have emerged as a promising solution for neural networks, allowing for energy-efficient computing units. The Politecnico di Milano has been developing programmable photonic processors integrated on small silicon microchips for data processing and transmission, which are now being used to construct photonic neural networks.
An artificial neuron, much like its biological counterpart, performs simple mathematical operations such as addition and multiplication. However, in a densely interconnected neural network, the energy cost of these operations increases exponentially, making them prohibitively expensive. To address this issue, the Politecnico di Milano has developed a chip with a photonic accelerator that utilizes a programmable grid of silicon interferometers to perform calculations quickly and efficiently. The time required for these calculations is less than a billionth of a second, making it an incredibly fast and energy-efficient option.
The photonic neural network has long been recognized as a promising technology, but the missing piece has been the ability to train the network. In this study, the Politecnico di Milano has successfully implemented training strategies for photonic neurons, resulting in a highly precise and fast-learning photonic “brain” that offers significant energy savings compared to conventional neural networks. This breakthrough holds great potential for artificial intelligence and quantum applications.
The device is not limited to neural networks but can also function as a computing unit for a range of applications, including graphics accelerators, mathematical coprocessors, data mining, cryptography, and quantum computers. The research is published in the prestigious journal Science.
Source: Polytechnic University of Milan