I am genuinely interested! Empirical research around how well our ad-hoc estimates (of gradients, variances, and Fisher information for example) perform is surprisingly limited, since it needs to be constantly reevaluated as SOTA changes

Should we be trying to detect distribution shift and correct it (eg. by taking more samples at the same set of parameters)? Does this matter in some models and not others?

The exponential decay mollifies the severity of distribution shifts, even sudden ones, over many iterations. But what nature do we expect this distribution shift to take?

Adam depends on the gradient distribution during training, which, as far as I know, we don't understand well? Here, adapted from the Adam paper, v_t is the var estimate, G_t is the gradient r.v. and X_t is an error r.v. for distribution shift.

Garner's usage dictionary calls these needless variants. Usage dictionaries outside of professional editing seem somewhat rare, but would be very useful in technical disciplines (if unrealistic to make.)

Probably not. I've generally found custom kernels to be less necessary when using JAX due to compilation (the benefit may be marginal compared to PyTorch.)

Screw it, here's a new JAX implementation in Optimsitix: github.com/packquickly/sched…

Neat! Heuristics based on this crop up in stochastic optimisation. Often in the argument of stepsize * sum/sqrt(squared_sum) being "roughly" bounded by stepsize. Of course the independent and symmetric assumptions are both violated but that never stopped anyone...

Ultimately I think the question is "who do we want getting into wser?" Optimising the lottery is not the hard part, it's agreeing on who we should be optimising for that is.

Finally, a million thanks to @PatrickKidger, who supervised this whole project. If you’re following me, chances are good you already follow him. If not, go give him a follow! (right after installing Lineax of course 😉) 4/4

The paper describes out how we achieved many of these things (such as differentiation through all our solvers,) and outlines some of the design choices we made when creating Lineax. arxiv.org/abs/2311.17283

Lineax offers numerous features over core JAX, such as: * stable gradients through least-squares solves; * compile times nearly twice as fast for iterative solvers; * a consistent API; * PyTree-valued vectors & operators; * structured matrices; * new solvers; and more! 💥

⭐ Lineax is now on arXiv! ⭐ If you’re doing linear solves or linear least-squares in JAX, give it a shot today! Lineax is fast ⚡️, has new solvers (eg. QR, tridiagonal), supports general linear Operators. github: github.com/google/lineax arXiv: arxiv.org/abs/2311.17283 1/n

depending on your philosophy, many would simply call this "mathematics" :)

They mentioned the choice between these two model functions made little difference in practice. While this is believable, and indeed proved to be true, it struck me as an example of a claim which is very difficult to verify using existing optimisation software. 12/12

This example was taken from an offhand comment in "Training Deep and Recurrent Networks with Hessian-Free Optimisation" by Martens and Sutskever.

This means that both algorithms made the same choice of when to increase/decrease the size of the trust-region, which is reasonable.