Urte Adomaityte

or dm me: happy to discuss, especially on exact asymptotics in ML and robust statistics!

- at the Maths for Modern ML (M3L) workshop, presenting arxiv.org/abs/2306.07104, spearheaded by @maha81193, F. Behrens, @MolecularRobot

Heading to #NeurIPS23! Come say hi: - at poster session 5 on Thu, presenting "Classification of Heavy-tailed Features in High Dimensions: a Superstatistical Approach" arxiv.org/abs/2304.02912…, work with supervisors @gabrielescr @PierpaoloVivo at @KCLDisorder

Another succesful PhD information event for our UG and MSc students at @kclmathematics under the umbrella of the @_Piscopia initiative 💅🏻 Big thanks to fellow PhD students L. Servius, A. Borkowski and @ShahpoOmar for their time!

🚨POSTDOCTORAL POSITION AVAILABLE🚨 in Complex Systems Modelling, Quantitative Modelling of Legal Complexity and Infonomics under my supervision in the [quantlaw.co.uk] Lab at @KCLDisorder @kclmathematics. Competitive salary, fixed term (27 months). Deadline 10th Jan 👇

my first submission to #NeurIPS2023 was accepted !! + additional results on Bayes-optimal performance and a separability threshold formula big thanks to @gabrielescr and @PierpaoloVivo ! @KCLDisorder @kclmathematics

The past two weeks were a blast in Cargèse at "Statistical Physics & Machine Learning Back Together Again" cargese2023.github.io. Around 100 of the top people in the field, including the next generation, discussed a lot of great science. I will miss you guys. We will be back!

@sebastiangoldt @gabrielescr @PierpaoloVivo @KCLDisSyst Thanks, see you soon!

@urteado @gabrielescr @PierpaoloVivo @KCLDisSyst Really neat work ! Looking forward to hearing more about it at Youth soon 🙂

fat tails alert!!! latest preprint on theory of machine leaning with @gabrielescr and @PierpaoloVivo at @KCLDisSyst : arxiv.org/abs/2304.02912 1/

What if labels are random? The training loss (bottom) is independent from the variance distribution, thus universal even for infinite variance (yellow)! No universality for test and train errors though. Thanks for reading! 9/9

Data separability/MLE-existence threshold? Computationally extracting with logistic loss: fatter tails -> more samples one can correctly classify; the threshold approaches known Gaussian data formula as a->inf. Infinite variance: Cover's result n_samples=2*dim is recovered 8/

And now published on @PhysRevE! journals.aps.org/pre/abstract/1…

(N/N) Bonus: mixed case of edges and 3-hyperedges — first-order phase transition persists but presence of a finite fraction of edges in the hypergraph makes the transition of second order. Thanks for reading!

(8/N) As coordination k of the hyperedges grows, we have easier recovery — the partial recovery phase shrinks. Also, the region in which it is information-theoretically impossible to fully reconstruct the signal (hard phase) rapidly shrinks to zero.

(1/N) First preprint alert! Here is a hyper-twist on planted matching problems for anyone interested in statistical physics methods for inference 🧵 In collaboration with @gabrielescr @zdeborova @toshniwal8 Link: arxiv.org/abs/2209.03423 @KCLDisSyst