Michael Burkhart

[Michael Burkhart @burkh4rt]

Update—the @uspto published our patent application for this work #today—https://t.co/avos2fNswg—I hope they don't actually work on #thanksgiving

[Michael Burkhart @burkh4rt]

Here, Gabriel Ruiz & I explore feature engineering for causal inference. We develop a neuroevolutionary approach that seeks representations least useful for predicting the treatment among those preserving as much information about the outcome as possible https://t.co/0tUeJxFmdI

[Michael Burkhart @burkh4rt]

Back in 2021, I was hired at @Adobe out of my phd by the extremely generous and talented Kourosh Modarresi. This past week, the @uspto awarded us #patent 11816562 for our work together on #recsys --
https://t.co/3mes5D1OZS
#HappyThanksgiving

[Michael Burkhart @burkh4rt]

For this and some results for the case that 𝑁 is not abelian, see our paper out today —
https://t.co/cIgExdlkOp
@CamOpenAccess @cambUP_maths

[Michael Burkhart @burkh4rt]

A finite group 𝐺 splits over an abelian subgroup 𝑁 iff for each prime 𝑝 a Sylow 𝑝-subgroup 𝑆 of 𝐺 splits over 𝑆∩𝑁 (Gaschütz). In this case, we can show that all such complements are conjugate iff for each prime 𝑝 all complements of 𝑆∩𝑁 are conjugate in 𝑆 (Cor. 1.3).

[Michael Burkhart @burkh4rt]

This follows from our Thm. 1.2 for finite groups: suppose 𝐽 acts on 𝑁 abelian via automorphisms & the induced semidirect product 𝑁⋊𝐽 acts 𝑁-transitively on some set Ω≠∅. If for each prime 𝑝, a Sylow 𝑝-subgroup of 𝐽 fixes an elt of Ω, then there is a 𝐽-invariant 𝜔∈Ω.

[Michael Burkhart @burkh4rt]

Our @iccs_conf paper on learning representations for causal inference was selected for a Journal of Computational Science special issue — check it out here:
https://t.co/QiwxRdTI4K
Gabriel & I thank the @iccs_conf for this honour

[Michael Burkhart @burkh4rt]

For details & proofs, see my new article ‘Conjugacy conditions for supersoluble complements of an abelian base and a fixed point result for non-coprime actions’ to appear in the Proceedings of the Edinburgh Mathematical Society
https://t.co/ZgAEFbH9s9
@Cambridge_Uni @cambUP_maths

[Michael Burkhart @burkh4rt]

Given a finite group 𝐺 with a normal subgroup 𝑁, 𝐺 splits over 𝑁 if there exists another subgroup 𝐻 such that 𝐺 is the semidirect product 𝑁⋊𝐻. In 1937, Zassenhaus showed that when the order & index of 𝑁 are coprime, 𝐻 exists (credit: Schur) & is unique up to conjugacy.

[Michael Burkhart @burkh4rt]

If for each prime 𝑝, a Sylow 𝑝-subgroup of 𝐻 fixes some element of Ω, then there exists an 𝐻-invariant element ω∈Ω.»

[Michael Burkhart @burkh4rt]

update—the @uspto granted us #patent 11,455,518 “User classification from data via deep segmentation for semi-supervised learning” for this work https://t.co/BrxigXRN0L

[Michael Burkhart @burkh4rt]

Kyle Shan & I explore a new approach to semi-supervised learning that iterates between refining a latent feature representation and performing low-density separation on said representation in our paper out this week—https://t.co/Ch2ukbORDI

[Michael Burkhart @burkh4rt]

If you take a subsampled version of Newton's method in #optimization and apply discriminative Bayesian filtering, you can derive a matrix-based version of Polyak's heavy ball #momentum @Cambridge_Uni @CambridgeC2D3 @CamOpenAccess Check it out here—https://t.co/Nh873y1xtR

[Michael Burkhart @burkh4rt]

It was a real pleasure working with Gabriel during his internship at @Adobe — for more thoughts on this work, tune into our talk at the International Conference on Computational Science (ICCS 2022) in London next month #MachineLearning