Optical neural networks more energy efficient than the human brain?
https://journals.aps.org/prx/abstract/10.1103/PhysRevX.9.021032
Also this YouTube video.
https://youtu.be/QD5D12mVJtw
However this guy is out of date about the flaw in current deep neural networks.
https://github.com/S6Regen/RP_AUTO_RESNET
The main advantages of optical neural networks are very low power requirements and high speed. The disadvantage is they are bulky with low data density.
If you can get say 1 million network evaluations per second then evolution algorithms will work as fast as back-propagation without all the hardware complexity.
If you can evolve a relatively small but powerful controller neural network connected to a large scale semiconductor associative memory system that could perhaps do very well as an intelligent system.
There is also this company who have an optical processing unit:
https://www.lighton.ai/our-technology/
They also have this paper on Direct Feedback Alignment training of neural networks: https://arxiv.org/abs/1906.04554v1
Iām not really a big fan of DFA training. I suspect it is using each layer in a relatively simplistic associative memory mode.