As the demand for faster, more efficient training of neural networks continues to grow, specialized photonic hardware has emerged as a potential alternative to classical computers for AI applications. Reservoir Computing (RC), a lightweight alternative to computationally-intensive Recurrent Neural Networks, has been demonstrated to be possible using simple delay dynamical systems. We propose an optoelectronic implementation of this architecture through a Mach-Zehnder modulator driven by delayed feedback from a laser. We introduce a new optoelectronic scheme in which input data is first pre-processed offline using two convolutional neural network layers with randomly initialized weights, generating a series of random feature maps. These random feature maps are then multiplied by a random mask matrix to generate input nodes, which are then passed to the reservoir computer. Such a scheme has achieved simulation results in-line with the state of the art for image recognition tasks, with a potential 10x increase in processing speed.
August 23, 2021
BSAC Project Materials (Current)
PREPUBLICATION DATA - ©University of California 2021