I'd never considered the islands-surrounded-by-fuzz before, but it makes me wonder a few questions - is there a programmatic way to identify the "islands" (without using a separate NN classifier)? Is trimming the "fuzz" a potential method of compression for models?
The dimensionality is a challenge here - these are 2D islands in a, what, 2000D space? Imagine a cross section of a cylinder, that would look like an island if it went through the cylinder, but not along it.
I wonder what it would look like if the cross-section plane was aligned with a line between two recognisable points.
When discussing “n parameters,” such as 6B parameters or in the case of GPT4 around 1.7T is the rumored parameters size, each parameter is a dimension. So, yeah, that’s a lot more than 2d or even 2000d. More like 1,700,000,000,000D.
The dimensionally of these islands must be insane. Every 2D representation is one of thousands upon thousands of possible slices through the weighting space.
Exploring these will be like trying to find the edge of fractals.
Any attempt at visibility into the inner workings of ML models should be welcomed IMO. It's going to be essential in the coming years, if we're going to reason about or regulate them. E.g. how would we hardcode Asimov's laws of robotics into some future deep learning AGI, if it's still just one big black box for us?
Thinking about law (As Asimov's are), I perceive essentially an impenetrable barrier in this possibility space. We create order from the chaos of potential behavior by defining limits.
Currently, in this 2000D StableDiffusion landscape, there are no boundaries for allowable "travel" and you quickly end up in the "sea".
So if we want AI to behave, we should research how to define the perimeter (in thousands / billions of dimensions) and "wall it off" so we can ensure the potentials inside the no-no space are inaccessible.
I'd never considered the islands-surrounded-by-fuzz before, but it makes me wonder a few questions - is there a programmatic way to identify the "islands" (without using a separate NN classifier)? Is trimming the "fuzz" a potential method of compression for models?