Project goals.
Project goals.
Embedded Files
The project aims to realize a hardware implementation of a neural network using polaritons, speeding up pattern recognition tasks to unprecedented performances and with reduced power consumption and rendering it compatible with standard electronics. The detailed project objectives are:
Design and optimization of optical processors considering fabrication limits and materials nonlinearities,
Proof-of-principle demonstration of a polariton-based photonic chip implementing image or sound recognition, integrated with laser diodes or large-scale μ-LED arrays,
Fabrication of a polaritonic neural network performing biomarkers detection for efficient genome analysis,
Integration of the photonic neural network operating at room temperature with a commercial ST microcontroller.
Long-term project vision.
Long-term project vision.
One of the approaching challenges to be overcome for the continuation of Moore’s law in current electronics is the von Neumann bottleneck, which arises due to the separation between processing and memory elements in standard information processing systems. And, even though information can travel back and forth at very high rates, these rates are finite and their limitation becomes more and more significant with each new generation of ever more powerful processors and larger memories. At the same time, the rapid rise of machine learning is providing a growing number of research fields with powerful tools for the recognition of patterns in data and for making predictions based on those patterns. Often machine learning techniques are built on neural network architectures, in which information is processed in parallel and there is no separation of memory and processing elements. While they are most often implemented in software, the possibility of implementing neural networks as hardware is seen as one route to overcoming the limitations of the von Neumann bottleneck. Moreover, such new technology, based on Artificial Intelligence (AI), would open different computation schemes beyond those based on binary data processing, leading to a paradigm shift in data analysis, modeling and cognitive computing. However, such promising technology has no real implementation in present day devices. In this project we propose, in an extraordinary synergetic effort of research institutes and universities across Europe, together with private companies, to implement a hardware-based photonic neuromorphic accelerator to be integrated into specific microcontrollers. Such an accelerator will be based on a hybrid electro-optical neural network capable of executing neuromorphic computation in an integrated device at ultra-high speed and with a strongly-reduced energy cost per operation.
We shall explore its applications in three areas: the recognition of images, sounds and genome-wide patterns of biomarkers. For the first two recognition tasks such a neuromorphic hardware accelerator will be able to increase the processing speed compared to present computation based on pure software simulation of convolutional neural networks. The third recognition application will radically simplify as well as substantially shorten the time required for genomic analysis based on microarrays, which are acquiring an ever-increasing role for clinical diagnostic/prognostic purposes and for drug development. The working principle of this device is based on a new observation made by a few groups involved in this project and published in the past two years: as theoretically predicted and experimentally demonstrated, a mixed state (so-called polariton) made of light and electronic excitations in a semiconductor–that can be described as a photon dressed with interactions–is able to outperform not only the linear classifier but also many other realizations of artificial neural networks that have been proposed using electronic or photonic technology. The main advantage that strongly interacting photons can offer for neuromorphic computation is their ability to work analogically, with strong nonlinearities inter- and intra-nodes and without the need for a costly optoelectronic conversion. Moreover, photons can propagate at ultrahigh speed and with very low energy consumption. These are fundamental elements, indispensable for the implementation of artificial neural networks able to surpass the standard von Neumann architecture of classical computers. This change of paradigm might have a profound impact on the field of smart autonomous sensors that, based on a series of signal inputs, need to make fast decisions. These devices already span a large range of applications that go from basic temperature, humidity or pressure measurements to more advanced image recognition in autonomous vehicles and market electronics, as well as to wearable sensors for health risk factors monitoring.
Only in 2019 the global sale of autonomous vehicles requiring such technology amounted to 1.4 million units, forecasted to attain about 60 million by 2028-2030, while the wearable sensors market in 2026 is expected to amount to 4.3 billion dollars. The three applications addressed in the project have been chosen to demonstrate, in representative cases, a potential advantage of our hardware-based neuromorphic accelerator over conventional technologies and achieve, thereby, a general technology acceptance. In turn, the technological development will feedback fundamental research, as networks with more complex interaction schemes and functionalities will be certainly searched for.
Graphics: Mateusz Król
Webpage design: Andrzej Opala
PolArt consortium:
PolArt consortium:
This project has received funding from the European Union's Horizon Europe EIC Pathfinder open action under the grant agreement No. 101130304 (PolArt).
Page updated
Google Sites
Report abuse