The exponential growth of generated data has led to the expansion of artificial intelligence (AI) for a broad range of applications, from medical diagnosis to autonomous driving. However, as the endemic processor-memory bottleneck of the traditional von Neumann architecture severely limits the efficiency of today’s processor, there is a current call for new computational paradigms that cope with the provisions for the big data era. The PHOTOART project arises in this scenario, on the path to accomplish the long-awaited photon-based computing, which has been envisioned to address the limitations of classical electrical computing platforms by offering improvements in terms of computational throughput and energy efficiency.
Thereby, the main pursue of the PHOTOART project will be the development of active devices on the silicon photonics platform for fostering the advance of next generation photonic hardware for AI applications. For this purpose, phase change materials (PCMs) and transparent conducting oxides (TCOs) have been chosen as advanced materials to be integrated on hybrid silicon waveguide structures, to develop the basic building blocks for integrated photonic neural network architectures. Such materials account for different properties: PCMs offer the possibility to implement long-term memory, with very low, ideally zero, energy consumption, while TCOs offer ultrafast modulation and strong non-linear effects. Therefore, key building blocks starting from the weighting banks to the non-linear activation functions will be developed.
All efforts will be focused on the achievement of coherent-type photonic neural networks. Principal innovations will include the optimization of a new type of amplitude and phase modulator, necessary to implement an emerging type of complex-valued neural networks. Such devices will also enable ultracompact wavelength-selective operation which, in principle, can be used to implement wavelength-division multiplexing (WDM) into the coherent scheme. On the other hand, the development of electro-optical phase shifters with non-volatile behaviour and the possibility to develop electrically tuneable and reconfigurable activation functions will also constitute important innovations of our project.
Finally, a demonstrator of an integrated coherent photonic neural network loaded with reconfigurable non-linear activation functions will be targeted. Such a demonstrator will be validated in an AI application. In short, the present PHOTOART project will aim to provide specialized hardware capable of fast and efficient neuromorphic computing acceleration for the big data era.