Technology: AMF Si-Photonics
Stefano Biasi, Riccardo Franchi, Lorenzo Pavesi
In search of a better performance
Photonic Integrated Circuits (PICs) have demonstrated state-of-the-art performance in challenging computational tasks, meeting the demand for scalability by mitigating power consumption.
Feed-Forward Neural Networks (FFNN) have been integrated in silicon photonics and have shown excellent performances compared to their electronic counterpart [1]. However, training these networks is powerhungry and time-consuming because of the many parameters to be optimized. Promising alternatives are neural networks that do not require full control of all the node interconnections, such as the Extreme Learning Machines (ELM) [2].
Optical image of the part of the chip that realizes the photonic ELM neural network. The blue lines indicate the optical paths, i.e. the waveguides and microresonators. The input and output gratings are highlighted by the yellow circles on the very left and very right, respectively. The input, hidden and readout layers, are highlighted by red, green and light blue contours, respectively.
Developing the photonic ELM
The design we present aims at a PIC which implements the computational paradigm of an ELM to get better performances than its electronic counterpart. The ELM is a FFNN consisting of a single hidden layer where training occurs only at the readout stage.
The created device contains an input layer where the information is encoded, a hidden layer formed by an array of 18 microresonators, and a final readout layer consisting of a linear classifier (LC). Thus, all the steps, from data encoding to network training, take place in the PIC.
The chip wire bonded and mounted on the PCB under the testing setup. The two metal holders for the input/output optical fibers are also observed.
Solving common benchmarks
A first proof-of-concept demonstration of the microresonatorbased ELM is reported in [3]. Common benchmarks are solved with high performances though with a low operational speed. The novel fully integrated PIC design is still under test while we anticipate better performances and operational speed.
Acknowledgements
This project has received funding from the European Research Council (ERC) (grant agreement No 788793, BACKUP), and from the MIUR (PRIN PELM 20177 PSCKT).
References
[1] Y. Shen, N. C. Harris, S. Skirlo, M. Prabhu, T. Baehr-Jones, M. Hochberg, X. Sun, S. Zhao, H. Larochelle, D. Englund, and M. Soljacîc, “Deep learning with coherent nanophotonic circuits,” Nat. Photonics 11(7), 441–446 (2017).
[2]G.-B. Huang, Q.-Y. Zhu, and C.-K. Siew, “Extreme learning machine: Theory and applications,” Neurocomputing 70(1–3), 489–501 (2006).
[3] S. Biasi, R. Franchi, L. Cerini, and L. Pavesi. “An array of microresonators as a Photonic Extreme Learning Machine,” APL Photonics 8, Sept, (2023).
