Francesco Orabona

As promised, I compiled all my lecture notes on Online Learning in a single PDF. Feedback is welcome! "A Modern Introduction to Online Learning" https://arxiv.org/abs/1912.13213  PS Happy New Year!

People of Milan, I will give a talk on December 19th at the University of Milan (via Celoria 18). I will talk about parameter-free machine learning through coin-betting

In our paper, we close the gap: we design a variant of Ridge Regression that matches the lower bound in *all* the cases. Also, we show novel rates in the case that the Bayes risk is 0. The proof is also interesting: we use online learning tools to prove it. 6/6 The end.

In this case, you can still achieve the performance of the best regression function, even if technically it is outside of your space! The rate will depend on how "infinite" is its norm and how "big" is the space. Papers by @lrntzrsc are the best reference on this topic. 4/6

Now, with kernels things are fuzzier. For example, Universal Kernels can approximate any continuous function. So, for universal kernels the optimal function can be always written in terms of the eigenfunctions of the kernel, but its norm might be infinite! 3/6

First, let's see what is the right way to study things with kernels. In finite dimensional space, the optimal regression function is in the space or outside of it. If inside, sooner or later you will converge to it. If if is outside, you cannot achieve its performance. 2/6

An important note: do you know that SVRG was invented independently by two groups at the same conference? Apparently ~98% of you is missing the second reference: https://papers.nips.cc/paper/4941-mixed-optimization-for-smooth-functions … Please do the right thing and always cite both papers!

After the lottery for the registration, apparently we need a lottery to see the poster session at @NeurIPSConf They shut down the doors and thousands of the people are out

ML and Optimization people, we have 2 positions for assistant professors in the ECE department at Boston University. We have strong research, friendly environment, and we are in a city CS+ECE is 13th on CSRankings for ML+CV: http://csrankings.org/#/index?vision&mlmining … http://bu.edu/eng/tenure-track-assistant-professor-engineering-applications-of-information-and-data-science-and-systems-engineering/ …