Correct me if im wrong, this is just a question originating from my own curiosity, but could you not apply the same or similar techniques to classical NN architectures? I dont see exactly how the differentiability of the architecture matters? I.e. in between epochs randomly perturb weight values (similar effect I understand to introducing scheduled noise or non-isomorphic transforms on the features), implement a scheduled causal drop in/out scheme etc? I know of people that train models these ways as a form of regularisation and they claim better fits too.
> I dont see exactly how the differentiability of the architecture matters?
I don't know that classical architectures are infeasible to perturb as noted in other comments. I just think we have a really interesting knob to turn in this specific case that doesn't seem to exist in other architectures.
How would you go about constraining the resource consumption (by up to half!) of a classical ANN while preserving realistic hopes that it will recover between training epochs? And, we are talking real resource reduction here not some academic thing. You can see the difference in task manager. That's actually the primary tool I've been using to determine how to tune the parameters of this scheme.
They have a fixed compute budget which is lower than what the LLM need, and dynamically decide to allocate this compute budget to different part of the network.
So its not exactly what you propose since here the compute budget is fixed (that's the point of the paper: to make the network learn how to allocate the resources by itself) but its dynamic for each part of the network, so it shows that its possible.
Enjoy your adventures in the Dark Forest.