Photonic probabilistic machine learning using quantum vacuum noise.

Area of Science:

• Quantum computing
• Machine learning
• Photonics

Background:

• Probabilistic machine learning relies on randomness for uncertainty encoding.
• Quantum vacuum noise offers a source of high-speed, energy-efficient randomness.
• Limited photonic hardware exists for controlling stochastic elements in probabilistic machine learning.

Purpose of the Study:

• To implement a photonic probabilistic computer using a novel photonic probabilistic neuron (PPN).
• To demonstrate the PPN's capability in solving probabilistic machine learning tasks.
• To propose a pathway for scalable, ultrafast, and energy-efficient all-optical probabilistic computing.

Main Methods:

• Implementation of a PPN using a bistable optical parametric oscillator (OPO) with vacuum-level injected bias fields.
• Programming a measurement-and-feedback loop for time-multiplexed PPNs with electronic processors (FPGA or GPU).
• Utilizing quantum vacuum noise as a random seed for encoding uncertainty and generating samples.

Main Results:

• Successful demonstration of probabilistic inference and image generation on MNIST handwritten digits.
• Encoding of classification uncertainty and probabilistic sample generation using quantum vacuum noise.
• Proposed an all-optical probabilistic computing platform with ~1 Gbps sampling rate and ~5 fJ/MAC energy consumption.

Conclusions:

• The developed photonic probabilistic computer offers a scalable, ultrafast, and energy-efficient hardware solution for machine learning.
• This work advances the integration of quantum phenomena with AI for next-generation computing.
• The proposed all-optical platform promises significant improvements in speed and energy efficiency for probabilistic computing.