Source-linked AI summary
Polaritonic neuromorphic computing outperforms linear classifiers
D. Ballarini, A. Gianfrate, R. Panico, A. Opala, S. Ghosh, L. Dominici, V. Ardizzone, M. De Giorgi, G. Lerario, G. Gigli, T. C. H. Liew, M. Matuszewski, D. Sanvitto
TL;DR
Earlier hardware reservoir computers had not surpassed linear classification on MNIST, motivating physical networks with faster, more efficient processing. This paper implements a reservoir computer using an exciton-polariton lattice and reports higher recognition than linear classification and previous hardware implementations.
Problem
Earlier hardware reservoir computers showed recognition rates below linear classification applied directly to MNIST, while physical neural networks are sought for faster, more efficient processing.
Source: p. 2Method
An 8 × 8 lattice of coupled exciton-polariton nodes processes digit inputs through nonlinear transmission, with a computer-trained linear readout classifying the output.
Source: pp. 2–3Results
93% recognition on MNIST exceeds previous hardware reservoir-computing implementations and is reported as the first result to surpass linear classification for pattern recognition.
Source: p. 3Takeaways & Limitations
Exciton-polariton interactions provide nonlinearities while the photonic component supports connectivity and high operational speeds.
Source: pp. 10–11The experiments use a continuous-wave pump; the authors anticipate that following transient dynamics could further improve recognition.
Source: p. 10
Abstract
from arXiv · showhide
Machine learning software applications are nowadays ubiquitous in many fields of science and society for their outstanding capability of solving computationally vast problems like the recognition of patterns and regularities in big datasets. One of the main goals of research is the realization of a physical neural network able to perform data processing in a much faster and energy-efficient way than the state-of-the-art technology. Here we show that lattices of exciton-polariton condensates accomplish neuromorphic computing using fast optical nonlinearities and with lower error rate than any previous hardware implementation. We demonstrate that our neural network significantly increases the recognition efficiency compared to the linear classification algorithms on one of the most widely used benchmarks, the MNIST problem, showing a concrete advantage from the integration of optical systems in reservoir computing architectures.