SPEAKER PROFILE |
 Dr Jack C. Gartside Blackett Laboratory, Department of Physics and London Centre for Nanotechnology. Imperial College London United Kingdom |
Retinomorphic Machine Vision in a Nonlinear Photonic Lasing Network
Abstract
As AI energy consumption and compute demand continue to skyrocket, developing new solutions for AI-specific hardware architectures is a pressing research direction. Neuromorphic computing systems, where complex physical dynamics directly implement AI processing, have emerged as an attractive candidate approach. A range of physical systems are being considered including magnetic, memristive and photonic - with photonic physics offering highly appealing benefits including high speed and bandwidth, and low energy consumption.
However, current photonic neuromorphic schemes typically operate in a linear regime, lacking crucial nonlinearity required for strong computing, and occupy large on-chip footprints. Here, we demonstrate an InP photonic network laser that behaves as an efficient neuromorphic machine vision platform [1], providing nonlinear processing in a compact 100 micron footprint. The network laser hosts many strongly-coupled, spatially-distributed lasing modes which operate in a “retinomorphic” manner inspired by retinal ganglia, performing nonlinear image processing through optical mode competition including parallel feature detection, classification of biomedical cancer images and tumour image segmentation.
We demonstrate reconfigurable physics-aware training of the nonlinear photonic dynamics, leading to tunable computation performance across single and multilayered network architectures. The performance of the system highlights the potential for highly nonlinear, spatially compact on-chip neuromorphic photonic systems.
[1] Ng, W.K., Dranczewski, J., Fischer, A., Raziman, T.V., Saxena, D., Farchy, T., Stenning, K., Peters, J., Schmid, H., Branford, W.R., Barahona, M., Moselund, K., Sapienza, R, & Gartside, J. 2024. Retinomorphic Machine Vision in a Network Laser. arXiv
However, current photonic neuromorphic schemes typically operate in a linear regime, lacking crucial nonlinearity required for strong computing, and occupy large on-chip footprints. Here, we demonstrate an InP photonic network laser that behaves as an efficient neuromorphic machine vision platform [1], providing nonlinear processing in a compact 100 micron footprint. The network laser hosts many strongly-coupled, spatially-distributed lasing modes which operate in a “retinomorphic” manner inspired by retinal ganglia, performing nonlinear image processing through optical mode competition including parallel feature detection, classification of biomedical cancer images and tumour image segmentation.
We demonstrate reconfigurable physics-aware training of the nonlinear photonic dynamics, leading to tunable computation performance across single and multilayered network architectures. The performance of the system highlights the potential for highly nonlinear, spatially compact on-chip neuromorphic photonic systems.
[1] Ng, W.K., Dranczewski, J., Fischer, A., Raziman, T.V., Saxena, D., Farchy, T., Stenning, K., Peters, J., Schmid, H., Branford, W.R., Barahona, M., Moselund, K., Sapienza, R, & Gartside, J. 2024. Retinomorphic Machine Vision in a Network Laser. arXiv
Bio
Jack C. Gartside is a Lecturer in Neuromorphic Computing, & leads the Neuromorphic Metamaterials group at Imperial College London. Jack & his group’s research interests cover neuromorphic computing architectures and algorithms, nanophotonics, nanomagnetism, magnonics, & optical magnetic switching.