My issue with #NeurIPS2025: if a co-author fails to submit their assigned reviews, jointly submitted papers are not receiving reviews.
Collective punishment is unfair and harmful,
discouraging collaboration, penalizing innocent researchers, worsening the academic culture.
We introduce a Hilbert space version of the Fuk-Nagaev bound to the integral operator framework and generalize the typical “Bernstein-trick”: we use the capacity of the hypothesis space as a proxy for the higher moments, giving optimal rates.
We show that the optimal rates from the work by Caponnetto & De Vito (2007) hold without the assumption of sub exponential noise - you only need some finite higher moment.
Check out our new result on regression with heavy-tailed noise !
I learned a lot on this project, thanks to @gaussianmeasure for leading the project.
@moskitos_bite@ArthurGretton
📢Next speaker in our online seminar about inverse problems and learning theory is
Dirk Lorenz @Dirque_L
with
** Learning regularizers - bilevel optimization or
unrolling? **
When: June 18, 14:00–15:00 (CET)
📢Next speaker in our online seminar about inverse problems and learning theory is
Dirk Lorenz @Dirque_L
with
** Learning regularizers - bilevel optimization or
unrolling? **
When: June 18, 14:00–15:00 (CET)
With another semester ending, here's your annual reminder that teaching evaluations systemically disadvantage women. Even when controlling for grades and other factors, students (esp. males) consistently give female professors lower scores. This can have serious ramifications.🧵
📢Our paper just got out:
For early stopped GD we derive minimax
rates of convergence. On our way, we precisely keep track of the number of hidden neurons required for generalization and improve over existing results.
sciencedirect.com/science/ar…