OSP

[bachatero]

How do you feel about the idea of OSP or Optical Signal Processing for sound systems?

I'm already working on a DSP (Digital Signal Processing) project here but there is a whole other world in optics waiting to leave the laboratory. The idea is to use optical components like semiconductor amplifiers and lasers to do tasks like convolution and FFTs at literal light speed, so we could get 0 latency perfection once it's developed.

I'm planning on researching this maybe for a PhD so I'm all ears!

 

[kemmler3D]

Sounds pretty interesting to me. If you could get these things to be effectively realtime, it would open up interesting possibilities. For example, maybe servo-control-like distortion control that works much higher in frequency...?

 

[BillH]

There are a few practical applications where optical processing is used. High pass and low pass filters in free-space optical systems being the simplest. In the past few years people have also shown diffractive optical (THZ) plates that form convolutional neural networks. There is also some work constructing matrix multipliers in the optical domain. One challenge with some of these later examples is the need for reasonably high photon fluxes which makes them difficult to apply to unamplified naturally occurring light fields in the environment.

 

[bachatero]

One challenge with some of these later examples is the need for reasonably high photon fluxes which makes them difficult to apply to unamplified naturally occurring light fields in the environment.

Hopefully this isn't as big of an issue as you might think because high power LDs are getting cheaper along with LEDs, and even if it costs thousands for a non-LD laser then that's cheaper than the alternative with DSP.

 

[BillH]

Hopefully this isn't as big of an issue as you might think because high power LDs are getting cheaper along with LEDs, and even if it costs thousands for a non-LD laser then that's cheaper than

For some applications you can add light- sure. I was commenting that for a fully passive optical system you often don’t have adequate flux for the losses in the processing steps and for conversion from optical to electrical for readout while also maintaining bit depth. As AI moves to lower bit depth models this concern is partially alleviated. For example, imagine you want to do all the processing of a Tesla collision avoidance system in the optical domain using the inherent photon fluxes on a cloudy day. Add to that the fact that the optical processing elements in the processor are narrowband. Consider the flux needed for the nonlinear element of the CNN that follows the matrix multiply and accumulate (MACs). The flux currently comes up short. I’ll try and follow up with a great plot comparing electronic processors to optical processors in terms of MACs per Joule and clock speed. It is very cool stuff.