Low-power scalable multilayer optoelectronic neural networks enabled with incoherent light
Author Information
Author(s): Alexander Song, Sai Nikhilesh Murty Kottapalli, Rahul Goyal, Bernhard Schölkopf, Peer Fischer
Primary Institution: Max Planck Institute for Medical Research, Heidelberg, Germany
Hypothesis
This study introduces a multilayer optoelectronic computing framework that alternates between optical and optoelectronic layers to implement matrix-vector multiplications and rectified linear functions.
Conclusion
The multilayer optoelectronic neural network achieved a recognition accuracy of 92% on the MNIST dataset and 86% on a nonlinear spiral classification task.
Supporting Evidence
- The multilayer optoelectronic neural network demonstrated a classification accuracy of 92.3% on the MNIST dataset.
- The system significantly reduces the number of read-ins and read-outs required for processing.
- Experimental results showed high correlation between simulated and actual outputs in the hidden layers.
Takeaway
This study shows how we can use light to help computers recognize images better and faster while using less energy.
Methodology
The study used a multilayer optoelectronic neural network with optical and electronic layers to perform matrix-vector multiplications and nonlinear activations.
Limitations
Variability in optical and electronic responses may limit performance, and the approach relies on analog computing.
Digital Object Identifier (DOI)
Want to read the original?
Access the complete publication on the publisher's website