Dust: Pretraining Transformers Without Backpropagation

api

It sounds like this is less computationally efficient than backprop, but more easily parallelizable. Is that fair?

polyomino

Even though this is way more expensive than backprop, could a hybrid approach where you fine tune an existing checkpoint that's been backpropped unlock further gains? It would be cool to apply this to different stages and see if that affects the learning trajectory

eriwang915

Dust's 243M model beating a 120x smaller one at most population sizes is the surprising part; bigger nets got more population-efficient, not less.

wrecked_em

Definitely more than meets the eye.

usernametaken29

> There are many interesting open questions. The first is whether, and how, Dust can find better directions than backprop’s first-order gradient Both algorithms are bound by the same Pareto frontier based on the Empirical Risk Minimisation Principle, so they’re already on the same trajectory. Interestingly backprop is limited by conditioning of the Hessian matrix in order to converge (differentiate correctly). So removing this limitation is actually a great step. I’m excited to see a comeback of evolutionary methods because they’re much more general, albeit costly and naive. We’re now very close to what can be described best as brute forcing the Pareto frontier out of our datasets. Not sure that’s what we want but I have no better ideas either.

oofbey

This is super yawn-worthy. Instead of backprop for the exact gradient you can run forward passes a thousand times with perturbed weights and get a Monte Carlo estimate of the gradient. Not very clever. Extremely NOT useful. But I guess the industry is littered with techniques for computing the same thing but vastly slower that some people find interesting. Homomorphic encryption. Zero knowledge proofs. Blockchain computing. Except in those cases there might be a legitimate reason to use it occasionally.

blt

Every few years, a derivative-free neural network optimization algorithm gets some hype. I'd bet my life savings that none of them ever make an impact. Derivative-free optimization can be useful for genuinely discontinuous objectives [1], but common neural network objectives are smooth and/or Lipschitz. The gradient is useful. Instead of trying random directions and hoping that one of them is an improvement, it tells you where to go. The more parameters you have, the more useful it becomes. A strict complexity gap between gradient-based and derivative-free Lipschitz convex optimization has been suspected for decades and recently proved (using AI, [2]). Neural net optimization is nonconvex, but not radically different. IMO, a more promising direction is gradient-based optimizers specialized to the neural network structure, like Muon [3]. [1] https://arxiv.org/abs/2202.00817 [2] https://arxiv.org/abs/2607.13335 [3] https://jeremybernste.in/writing/deriving-muon

wg0

It's computationally expensive and infeasible but what's the upside? Genuine question due to unfamiliarity with the subject.

AIorNot

Hmm I was more interested in this alternative to backprop: https://news.ycombinator.com/item?id=49701182

syntacticsalt

I'm skeptical as to whether zeroth-order methods really lend themselves to a Bitter Lesson argument. First-order methods don't explore the loss landscape optimally, but the loss function tends to be nonconvex, and zeroth-order methods don't address that issue head on. Dust smooths, and so do applicable first-order methods. Remove the nonconvexity issue, and I suspect Dust's purported advantages evaporate (based on published theoretical work), so it's pretty odd to me that the paper never discusses convexity. I could buy that this method scales better than previous zeroth-order methods, and that's interesting, but it doesn't seem like enough of a moat to keep improved first-order methods from drinking its milkshake, except in cases where a zeroth-order method is already a primary option: the network needs to call a simulator that doesn't expose gradient-like information. (In cases where gradients don't exist , I'd still argue for other options, e.g., Clarke-generalized gradients where applicable, so long as those can be computed with the available information. I know this technology has been published for automatic differentiation, so I would imagine it could be incorporated into backprop and used with a suitable optimization algorithm.)