Ivan Boikov (Thales)
Ivan’s work in the POST-DIGITAL project culminated in two major results, both aiming for scalable on-chip all-optical neural networks.
The first contribution was towards scalability improvement of on-chip neural networks in the reservoir computer architecture. There, Ivan has proposed to use evanescent coupling between nonlinear resonators, which allowed to drastically reduce chip footprint of the internal layer of reservoir. He performed an extensive analysis of such a reservoir and linked how basic parameters of resonators and their coupling contributes to computing performance in a very general sense. Ivan also has proven that such a reservoir is capable of solving nontrivial tasks in real-time (such as recovery of nonlinear distortion in an optical fiber) and, depending on a material, a bandwidth of up to tens of gigahertz. These results have been published in New Journal of Physics [1] and presented at CLEO 2023 [2] and MRS Spring Meeting 2023 [3]. A more recent work concentrates on highly relevant problem of nonlinear channel equalization in passive optical networks (PON), available on arXiv [4]. Ivan has also designed a layout for a chip prototype that has been fabricated successfully, and preparations for the experiments are currently being carried out.
In the last part of his Ph.D., Ivan concentrated on another work that tackled computing based on spikes. He has considered the recently demonstrated two-section photonic crystal nanolasers for stochastic neuromorphic computing (PhD work by M. Delmulle between Thales and C2N). Such a laser is not only compact, but also allows coupling to waveguides, which is favourable for creating scalable neural networks. Depending on pumping strength, this laser can emit optical spikes with random intervals comparable to a nanosecond. The theoretical analysis Ivan conducted has shown that such lasers could accurately sample from a Boltzmann machine, which allows for stochastic inference at considerably faster timescales than possible with electronic hardware. Recent simulations show that such behavior scales to large networks of 100s of nanolasers and can therefore perform nontrivial computation. A part of results are included in Ivan’s Ph.D. thesis, but not published yet.
As Ivan’s work at Thales continues due to a 4-year PhD track, he plans to:
- publish additional results on integrated reservoir computer,
- finally carry out experiments on the chip prototype,
- extend and publish the work on the spiking nanolaser.
[1] https://iopscience.iop.org/article/10.1088/1367-2630/acfba6/meta (DOI 10.1088/1367-2630/acfba6)
[2] https://ieeexplore.ieee.org/abstract/document/10232521
[3] https://www.mrs.org/meetings-events/spring-meetings-exhibits/2023-mrs-spring-meeting/call-for-papers/presentations/detail/2023_mrs_spring_meeting/2023_mrs_spring_meeting-3836089
[4] https://arxiv.org/abs/2405.06102 (DOI 10.48550/arXiv.2405.06102)