The Work
Chen took on a hardware that cannot be trained conventionally: a nonlinear photonic network where no gradient passes through the medium, leaving only the input light pattern to learn.
What Worked
Forward-only contrastive algorithm running end to end on the device and substantially improved classification accuracy. The method also generalised into a family of single-pass continual-learning losses that beat the online state of the art.
What Has Been Achieved
Carrying a forward-only contrastive algorithm onto the instrument, the team built a gradient-free encoding pipeline that runs end to end on the device and substantially improved classification accuracy by changing only the input, generalised the method into a family of single-pass continual-learning losses that beat the online state of the art.
What Was Learned
The calibrated twin predicts the optics but cannot yet explain which input patterns the physics can act on, the limit that reframed the year's central question.
About
Xing holds a PhD in Physics and Electronics Engineering from Beihang University and Paris-Saclay University. She spent her time as a Post-Doctoral Researcher working on local, bio-plausible learning rules as an alternative to back propagation for neuromorphic hardware.
Lab & Advisors
Riccardo Sapienza
Imperial College London
Professor, Physics
Riccardo Sapienza is a Professor of Physics at Imperial College London specializing in photonics and laser technologies, leading groundbreaking research in controlling complex laser actions and their applications in neuromorphic computing.
Jack Gartside
Lecturer (Assistant Professor), Physics, Imperial College London
Jack Gartside is Assistant Professor at Imperial College London and PI of the Neuromorphic Metamaterials research group. Jack and his team focus on developing physics-based neuromorphic computing architectures and algorithms, exploring emergent complex physical dynamics in nanoscale photonic, magnetic, and semiconductor metamaterials to build efficient bio-inspired intelligent systems.
