Spatio-Temporal Photonic Liquid State and Extreme Learning Machines
Abstract
In this work, we proposed the Bayesian optimization algorithm for tuning the hyper-parameters in large-scale photonic reservoir computers. We tested this approach on a previously reported experimental system, applied to a challenging task in computer vision where 91.3% and 99% accuracy in image recognition of video clips from standard image recognition databases KTH and MNIST used to verify the performance of the developed Photonic Recurrent Neural Network (RNN). We also compared the results with the grid search and Bayesian optimization of the Photonic RNN, commonly used by the non-photonic RNN Computational (RC) community. We report improvements in terms of (1) the classification performance , with an improvement in accuracy of up to 4%, and (2) the convergence time to the optimal set of hyper-parameters, with a roughly 30% reduction in time (that could be doubled for a less than 1.5 accuracy penalty). Taking into account the proximity of the accuracy of our photonic reservoir computer to the state-of-the-art results on this task, and the experimental hyper-parameters optimization time measured in days, these improvements prove to be precious enhancements of the system performance. Furthermore, extensive exploration of the hyper-parameters space with the Bayesian method offers valuable insights on its underlying structure and the relative importance of the parameters. Considering all the advantages offered by the Bayesian optimization algorithm, it may soon become the new standard approach for the optimization of hyper-parameters in photonic reservoir computing.
Document Details
- Document Type
- Technical Report
- Publication Date
- Mar 01, 2022
- Accession Number
- AD1165520
Entities
People
- Damien Rontani
Organizations
- CentraleSupélec