• Citer

Frequency Multiplexed Photonic Reservoir Computers and Extreme Learning Machines

par Lupo, Alessandro ;Massar, Serge
Référence Physics and Simulation of Optoelectronic Devices(XXXIII: March 2025), Proc. SPIE 13360, Physics and Simulation of Optoelectronic Devices XXXIII, 133600G
Publication Publié, 2025

Publication dans des actes

• Accès en ligne
• Détails
• Contenu
• Statistiques

Résumé :

We present how the frequency degree of freedom of light can be used for neuromorphic computing. Since multiple frequencies can propagate in the same device without interacting, this provides a simple and natural way to exploit the parallelism of photonics in computing. In the works presented the information is encoded in the amplitudes of a frequency comb. Interference between comb lines is realized using electro-optic modulators or the Kerr nonlinearity in an optical fiber. This approach is demonstrated on two rather simple machine learning algorithms, Reservoir Computers and Extreme Learning Machines. These algorithms are well suited for preliminary implementations of neuromorphic computing because of the small number of parameters and interconnection weights that need to be trained. The Extreme Learning Machines uses a simple fiber optics feedforward architecture. Modifying this to a recurrent architecture that adds memory to the system, we implement a Reservoir Computer. A programmable spectral filter is used to implement the output weights in the optical domain. In this way two Reservoir Computers can be combined in a deep architecture, with the interconnection between reservoirs implemented entirely in the analog domain. We discuss the relation to other photonic neuromorphic architectures that exploit the frequency degree of freedom, but in a highly nonlinear regime (typically using femtosecond pulses), as well as the potential for implementation in integrated optics.