Quantum-limited stochastic optical neural networks operating at a few quanta per activation
2025

Quantum-limited optical neural networks using single photons

Sample size: 10000 publication 10 minutes Evidence: high

Author Information

Author(s): Ma Shi-Yuan, Wang Tianyu, Laydevant Jérémie, Wright Logan G., McMahon Peter L.

Primary Institution: Cornell University

Hypothesis

Can optical neural networks operate effectively in a regime where each neuron is activated by just a single photon?

Conclusion

The study demonstrates that high classification accuracy can be achieved in optical neural networks even when operating under extremely low signal-to-noise ratios.

Supporting Evidence

  • High classification accuracy of 98% was achieved on the MNIST dataset.
  • The optical energy used per classification was just 0.038 photons per multiply-accumulate operation.
  • Physics-aware stochastic training allowed for effective operation in a low signal-to-noise regime.

Takeaway

This research shows that we can use very few photons to make accurate predictions with neural networks, which is like using just a tiny bit of light to see clearly.

Methodology

The study involved training a single-photon-detection neural network (SPDNN) using a physics-based probabilistic model to handle the stochastic nature of photon detection.

Potential Biases

Potential biases may arise from the specific hardware used and the assumptions made in the probabilistic model.

Limitations

The experiments were limited to specific optical setups and may not generalize to all types of optical neural networks.

Statistical Information

P-Value

p<0.05

Statistical Significance

p<0.05

Digital Object Identifier (DOI)

10.1038/s41467-024-55220-y

Want to read the original?

Access the complete publication on the publisher's website

View Original Publication