Photonic probabilistic machine learning using quantum vacuum noise
Name
choi-seouc130-sm-eecs-2024-thesis.pdf
Description
Thesis PDF
Size
9.35 MB
Format
Adobe PDF
Checksum (MD5)
21b6004806625b5db311ee1ce26fc9f4
Author(s)
Choi, Seou
Advisor(s)
Soljačić, Marin
Date Issued
May 2024
Publisher
Massachusetts Institute of Technology
Abstract
Probabilistic machine learning is an emerging paradigm which harnesses controllable random sources to encode uncertainty and enable statistical modeling. The pure randomness of quantum vacuum noise, fluctuation of electromagnetic fields even in the absence of a photon, has been utilized for high speed and energy-efficient stochastic photonic elements. Nevertheless, the experimental demonstration of photonic probabilistic computing hardware has remained elusive so far, due to the lack of programmable stochastic optical elements which can implement probabilistic machine learning algorithms. Here, we implement a photonic probabilistic computer consisting of a programmable stochastic photonic element, which we refer to as a photonic probabilistic neuron (PPN). We implement this PPN using a biased optical parametric oscillator, which utilizes quantum vacuum noise to generate a tunable probability distribution controlled by a bias field. We then implement a measurement-and feedback scheme for time-multiplexed PPNs in electronic processors (FPGA or GPU) to solve certain probabilistic machine learning tasks. We showcase how we can encode probabilistic behavior in two representative models of machine learning, discriminative and generative models, by showcasing probabilistic inference and image generation of MNIST-handwritten digits. While solving these probabilistic machine learning tasks, quantum vacuum noise works as a random source which can encode classification uncertainty in inference and enable probabilistic generation of samples. Furthermore, we propose a path toward an all-optical probabilistic computing platform. We estimate the sampling rate of the PPN as ∼ 1 Gbps and energy consumption as ∼ 5 fJ/MAC. Our work paves the way for scalable, ultrafast, and energy-efficient probabilistic machine learning hardware.
MIT Department
Massachusetts Institute of Technology. Department of Electrical Engineering and Computer Science
Terms of Use
In Copyright - Educational Use Permitted
Copyright retained by author(s)
Persistent DSpace Link
