Profiling question

bigcupholder

Groupie
Messages
79
If you took two QCs and iteratively used one to profile the profile of itself on the other unit, say starting from one of the amp models, would this iterative recapture process eventually devolve into something completely different than the original model, does it ever stabilize at some other tone and if so, how many of these stable points exist?

Anyone have two QCs and want to try this out?
 
If you took two QCs and iteratively used one to profile the profile of itself on the other unit, say starting from one of the amp models, would this iterative recapture process eventually devolve into something completely different than the original model, does it ever stabilize at some other tone and if so, how many of these stable points exist?

Anyone have two QCs and want to try this out?

It'll be just like pointing a video camera at the tv.
 
Kempers would also work for this although I suspect their model is more constrained so the result would be less interesting.

It'll be just like pointing a video camera at the tv.
To clarify, I'm not suggesting running the captures simultaneously in case that was what you were thinking. This is the process:

1. Set up an amp model on unit A
2. Use unit B to capture unit A
3. Use unit A to capture unit B
4. Repeat steps 2 and 3 multiple times (say >10 times? I have no idea how long this takes)

If the profiling is really good, it should be stable with respect to itself.
 
Dearest Mother,

Day 237. My dual QC captures of captures no longer resemble anything approaching my beloved Peavey Iconic amp. It’s nothing more than audible rhythmic beeps. I believe it’s Morse code they are using to communicate with each other. I’ve experience weird anomalies in the house. Lights flickering. Radios turning on. My Cioks power adapter is missing. I believe they plan to kill me. Send help.

Regards,

Frederick
 
If you took two QCs and iteratively used one to profile the profile of itself on the other unit, say starting from one of the amp models, would this iterative recapture process eventually devolve into something completely different than the original model, does it ever stabilize at some other tone and if so, how many of these stable points exist?

Anyone have two QCs and want to try this out?
inception GIF
 
Kempers would also work for this although I suspect their model is more constrained so the result would be less interesting.


To clarify, I'm not suggesting running the captures simultaneously in case that was what you were thinking. This is the process:

1. Set up an amp model on unit A
2. Use unit B to capture unit A
3. Use unit A to capture unit B
4. Repeat steps 2 and 3 multiple times (say >10 times? I have no idea how long this takes)

If the profiling is really good, it should be stable with respect to itself.

I was just kidding.

I think someone did this with the Kemper already - you might search on TGP to find it.
 
If you took two QCs and iteratively used one to profile the profile of itself on the other unit, say starting from one of the amp models, would this iterative recapture process eventually devolve into something completely different than the original model

Yes. Non-literarities add up pretty quickly - and this has nothing to do with the quality of NDSP's captures. Information loss is inherent to the process; you just can't reproduce the source perfectly from a capture/profile.

The "camera recording a TV", or re-encoding the same video file over an over are good analogies. For an audible demonstration, check the last 7 minutes of NIN's The Background World.

 
Last edited:
Yes. Non-literarities add up pretty quickly - and this has nothing to do with the quality of NDSP's captures. Information loss is inherent to the process; you just can't reproduce the source perfectly from a capture/profile.

The "camera recording a TV", or re-encoding the same video file over an over are good analogies. For an audible demonstration, check the last 7 minutes of NIN's The Background World.


It's really not about non-linearities at all.

Cycle consistency is a pretty common concept in machine learning. For example, if your model translates some text from English to German and then back again, you should get the same phrase semantically. Neural DSP may have considered this and it may be totally stable in repeatedly capturing captures of itself. Or not, and it might produce weird and interesting tones.

The camera recording the TV analogy isn't quite applicable here because there's a lot more sensor noise in that process.
 
It's really not about non-linearities at all. Cycle consistency is a pretty common concept in machine learning.

Yeah, and so is cycle consistency loss. If you have a neural network setup converting, say, and image A to image B, the reverse process will yield something that can look a lot like image A - but will never be exactly the same.

The camera recording the TV analogy isn't quite applicable here because there's a lot more sensor noise in that process.

You have a chain of A/D, D/A, filtering, digital signal processing and ML networks happening within captures/profiles. All of them introduce noise, losses, or both.
 
Last edited:
Yeah, and so is cycle consistency loss. If you have a neural network setup converting, say, and image A to image B, the reverse process will yield something that can look a lot like image A - but will never be exactly the same.
Yeah I didn't want to get overly technical and confuse people by saying "loss". We're talking about the same thing.

If the model is highly constrained, say adjusting the levels on a fixed EQ and the threshold on some saturating non-linearity (repeated for depth), this could be very stable. Note that the cycle consistency here is with respect to the model parameters, not the signal. So signal noise is potentially irrelevant if the SNR is high enough or there's some redundancy in the capture that makes it robust to additive noise.

You have a chain of A/D, D/A, filtering, digital signal processing and ML networks happening within captures/profiles. All of them introduce noise, losses, or both.
Hence the curiosity. My guess is that the quantization noise and line noise are insignificant in the capturing process and any instability in iterative recapturing would come from the model itself.

P.S. are you also at NeurIPS?
 
Back
Top