Optical Neural Network Hardware
Problem being addressed
Neural network architectures have become increasing prevalent, but the question of what hardware works best to implement them is an open one. How can optical circuits be used to build neural networks?
Neural networks are the most commonly used tool in the AI community. However, implementing neural networks on standard hardware can be computationally expensive, and also energy expensive. This motivates the need for specialist hardware that is tailored specifically for the implementation of neural networks. In this paper a nanophotonic circuit is proposed which uses optics (programmable Mach-Zender interferometers) to implement a neural network hardware architecture. In particular, a two layer network optimized for vowel recognition was implemented and tested successfully. Further, it is proposed that optical systems may be far more effective than conventional hardware for implementing neural networks.
Advantages of this solution
This research provides a novel hardware implementation of neural networks. It's anticipated that optical computing hardware may provide faster hardware implementations for machine learning algorithms.
Solution originally applied in these industries
Possible New Application of the Work
Electronics and Sensors Industry
Since optical systems tend to be less energy intensive than standard computing, this type of hardware may find a place in IoT and other devices where energy is a concern, but AI is a requirement.
In the future we could see large scale optical neural network hardware built as a remote resource, thus speeding up the process of training machine learning data sets and opening up the doors for more complex, computationally intensive, problems to be tackled.
Source DOI: #############
Source URL: #############