Fancy displays on products - Yay or Nay?

Hate bending over to squint at little screens, don't like spending more on a device just for the screen, don't like the screen sucking a lot of resources from the device, don't like screens that are easily smashed, won't buy any device that does not have a fully functional MacOS App to edit on a full sized monitor with on-device screen/knobs only used for quick tweaks and/or backup editing if App is inoperable for some reason.

Other than that, those screens are wonderful.

Can't get my head around the whole neural network thing with the Quad Cortex - sounds like some marketing to me, but, if somehow it can understand what my brain wants to play and makes my hands play it in the tone my brain wants to hear it, then I'm all in.
 
Furthermore the computational power of the DSPs used is simply not great enough to run a WaveNet or RNN with any decent level of accuracy.
And that’s to run inference. Training is way more computationally expensive still, and takes a ton of training data, not something you could even begin to do in the context of a limited-compute floor device. There are approaches that rely on using very few training samples, or even just one, but that part of the field is still in its infancy and the results are, shall we say, only academic, and not very impressive even then (except perhaps in facial recognition, where best few-shot systems have been exceeding human accuracy for years now).

Also, as you (and the paper) mention, WaveNet is only impressive if you hand-pick examples. You kind of never know ahead of time whether it’ll be any good. It’s pretty much style transfer for audio - we’ve all seen and created those “stained glass” and Van Gogh portrait stylizations using phone apps, but half the time the result is garbage, and the remaining half is not the actual Van Gogh. 😂

My guess is, they either deliberately conflate the use of convolutions with “AI”, or use gradient descent somewhere to tone match presets and call that “AI”. To optimize an amp model in a real AI fashion end to end it’d need to be fully differentiable (and preferably also at least mostly convex so you don’t get stuck in local minima), you’d need lots of clean training data, and the end result could still end up not too hot.

This will work eventually. An iPhone has crazy levels of compute when it comes to neural nets. Eventually neural acceleration will filter down to the chips one can actually buy and build products with, and people will nail down the algorithmic side of things better too (it probably still won’t be perfect, but for guitar you don’t need perfect, you need weird and usable). I’d say optimistically within 5 years we might start seeing something. But Cortex ain’t it.
 
That is what I have suspected. I'm not an expert but I've been studying AI, neural nets, etc. the last few months and the claims don't hold water.

Furthermore the computational power of the DSPs used is simply not great enough to run a WaveNet or RNN with any decent level of accuracy. Their own paper claims only 75% success rate in blind A/B tests using a WaveNet1 model which only runs at 1.9x real-time on a 2.5 GHz computer. There's no way that's going to run in real time on a 500 MHz processor. To get better results requires a WaveNet3 model which barely runs in real-time (1.1x) at 2.5 GHz. I suspect the test results would be worse using guitar players that are familiar with overdriven guitar tone. A clarinet player isn't exactly the target user.

Additionally the aliasing performance is terrible. Not just below average but downright terrible. In their tests the model had aliasing components that were GREATER in amplitude than the harmonics.

My personal guess is that they just use NN techniques offline to fit nonlinear transfer functions of whatever device they try to model. The drawback of this method is that the result doesn’t give any relation to the physics of the device (amp, effects, etc...). At least that’s what I know from some advanced control theory lectures, when it came to NN in signal processing. I neither think that they use NN for real time processing.
 
Thing is, if you have the schematic, there's nothing to fit really. You already have all the transfer functions. And if you don't have the schematics, then a high gain amp will be darn near impossible to fit well because nonlinearities are so severe. Note the emphasis on well. Neural net is provably an universal approximator, and as such it can theoretically approximate anything to an arbitrary degree of accuracy. This is mathematically proven. It just might not be possible to train it, and it might require you to boil the ocean in order to run. And we should be thankful for it, because that's why we don't have Skynet.
 
The touchscreen on the Gigboard is pretty cool and works well. Very quick and easy to add/make changes.

I bought the Gigboard exclusively to try the big touchscreen, turn out its a super fun unit, the looper alone has helped me store my ideas on the fly far better than anything else besides my phone.

The touchscreen has made me realise that i love to interact with this kind of units this way, cuts the adjustment times in such a significant way that going back the "old" way feels cumbersome.

Cant wait for a Fractal unit to implemente a touchscreen.
 
That is what I have suspected. I'm not an expert but I've been studying AI, neural nets, etc. the last few months and the claims don't hold water.

Furthermore the computational power of the DSPs used is simply not great enough to run a WaveNet or RNN with any decent level of accuracy. Their own paper claims only 75% success rate in blind A/B tests using a WaveNet1 model which only runs at 1.9x real-time on a 2.5 GHz computer. There's no way that's going to run in real time on a 500 MHz processor. To get better results requires a WaveNet3 model which barely runs in real-time (1.1x) at 2.5 GHz. I suspect the test results would be worse using guitar players that are familiar with overdriven guitar tone. A clarinet player isn't exactly the target user.

Additionally the aliasing performance is terrible. Not just below average but downright terrible. In their tests the model had aliasing components that were GREATER in amplitude than the harmonics.
I’m 100% ok with the aliasing which is terrible on their chart... only for simple sine fonction wave!!! What would it be for much more Complex input (Am I right?)
BTW on their website they show a quadcore 2GHz sharc dsp... not a 500MHz... so maybe could this allow over sampling like Doug Castro said. The fact is that I suffer from this aliasing on their plugins but don’t have the technical knowledge to make any measurements like they show on their wavenet paper... I know you can do that like easily.
or could your explain how I could measure it by myself?
thanks.
You must know that capture/profiling approach is very assuring for a lot of people. Having to wait for very hard work from you guys to get a new sound can be a little frustrating for some. I don’t care, with IR I can achieve pretty all the sounds I could imagine. (With the SOLO100, most of the time)
 
Thing is, if you have the schematic, there's nothing to fit really. You already have all the transfer functions. And if you don't have the schematics, then a high gain amp will be darn near impossible to fit well because nonlinearities are so severe. Note the emphasis on well. Neural net is provably an universal approximator, and as such it can theoretically approximate anything to an arbitrary degree of accuracy. This is mathematically proven. It just might not be possible to train it, and it might require you to boil the ocean in order to run. And we should be thankful for it, because that's why we don't have Skynet.
Isaac Arthur said it best,

With artificial intelligence,
Keep it simple
Keep it dump
Or else you'll end
Under Skynet's thumb
 
For the OP. I like a bigger screen. But it's somewhere like 5th on my list of priorities.

-Functionality (how well it does what it's supposed to do.)
-Usability (Do I need to be an EE or corresponding degree for what we're talking about to get the most out of it.)
-Durability. Need to know that I won't only get a year or 2 out of it.
-Support. Is this a 'one and done' deal?
-Pretty pictures or screen.
 
I think the axe fx 3 would be a lot better if it was blue tooth capable and have an app for your phone so you don't need to tote around a computer or have to monkey with the axe fx it's self and add a place for memory extension like a portable hard drive to be able to call up presets
 
Why would anyone need even more presets though? I use maybe half a dozen. I imagine a pro who’s constantly working on albums might need a hundred, maybe. After that you just can’t remember what you already have in there.

Agree on a tablet, wireless version of AxeEdit. That’d be pretty handy. Bluetooth strikes me as a bad choice for this, however - bandwidth and range are both pretty limited. FWIW UA OX uses WiFi.
 
Why would anyone need even more presets though? I use maybe half a dozen. I imagine a pro who’s constantly working on albums might need a hundred, maybe. After that you just can’t remember what you already have in there.

Agree on a tablet, wireless version of AxeEdit. That’d be pretty handy. Bluetooth strikes me as a bad choice for this, however - bandwidth and range are both pretty limited. FWIW UA OX uses WiFi.
There are more than a half a dozen bands out there with different tones. I like variety and easy access to those variety like from Eddie van Halen to Warren Demartini To BB King to Whom ever. I would much rather have it than not need it than need it and not have it. Makes since ?
 
for me, must have a screen. But I do not need touchscreen but would be OK with it.I'm fine with the screen on the FM3 because use FM3 Edit 99% of time.

BUT, ******* I still long for the day when ALL things FM3 are accessible\editable from FM3 Edit.
 
I think the axe fx 3 would be a lot better if it was blue tooth capable and have an app for your phone so you don't need to tote around a computer or have to monkey with the axe fx it's self and add a place for memory extension like a portable hard drive to be able to call up presets
FracPad works!
 
Why would anyone need even more presets though? I use maybe half a dozen. I imagine a pro who’s constantly working on albums might need a hundred, maybe. After that you just can’t remember what you already have in there.

Agree on a tablet, wireless version of AxeEdit. That’d be pretty handy. Bluetooth strikes me as a bad choice for this, however - bandwidth and range are both pretty limited. FWIW UA OX uses WiFi.
Not all of us are minimalists. Some of us create presets for every song because we like to create complex chains. There is nothing wrong with with either end of the spectrum, but if you're a minimalist how is having the ability to create hundreds of additional presets gonna hurt you? If you're into everything but the kitchen sink then yes, having only a limited preset storage space will negatively influence your buying decisions. And yes, Fractal should come with wireless ways to connect your wireless device.
 
Back
Top Bottom