Maria Refinetti
@mariarefinetti
ID: 1217064561712431105
14-01-2020 12:43:06
40 Tweet
127 Takipçi
76 Takip Edilen
In summer 2020, despite the pandemic, I and Krzakala Florent organized an in-person 2-week school in Les Houches. A highlight there was an incredibly productive open problem session and subsequent working groups. Two of them produced papers that are among my favourite in the past year:
[#FridayWiMLDSPaper 📜 curated by the WiMLDS Paris team] "Align, then memorise: the dynamics of learning with feedback alignment" by Maria Refinetti, Stéphane d'Ascoli, Ruben Ohana & Sebastian Goldt arxiv.org/pdf/2011.12428… #WiMLDS #WiMLDSParis
![[#FridayWiMLDSPaper 📜 curated by the <a href="/WiMLDS_Paris/">WiMLDS Paris</a> team]
"Align, then memorise: the dynamics of learning with feedback alignment" by <a href="/MariaRefinetti/">Maria Refinetti</a>, <a href="/stephanedascoli/">Stéphane d'Ascoli</a>, <a href="/oharub/">Ruben Ohana</a> & <a href="/sebastiangoldt/">Sebastian Goldt</a>
arxiv.org/pdf/2011.12428…
#WiMLDS #WiMLDSParis](https://pbs.twimg.com/media/FCSeN4rXoAQvk9d.jpg "WiMLDS Paris (@wimlds_paris) on Twitter photo [#FridayWiMLDSPaper 📜 curated by the <a href=\"/WiMLDS_Paris/\">WiMLDS Paris</a> team]
\"Align, then memorise: the dynamics of learning with feedback alignment\" by <a href=\"/MariaRefinetti/\">Maria Refinetti</a>, <a href=\"/stephanedascoli/\">Stéphane d'Ascoli</a>, <a href=\"/oharub/\">Ruben Ohana</a> & <a href=\"/sebastiangoldt/\">Sebastian Goldt</a>
arxiv.org/pdf/2011.12428…
#WiMLDS #WiMLDSParis")
S'il y a une exposition à ne pas rater cet hiver, c'est celle sur le peintre russe Ilya Répine au Petit Palais ! 😍🎨🖼️ oursmagazine.fr/exposition/lex…
Linear autoencoders perform PCA. What happens if you add a non-linearity? What representations do they learn, what's their performance, and... can we analyse learning on realistic data? 🤔 Happy to share this #preprint with Maria Refinetti: arxiv.org/abs/2201.02115 🧵1/3
📜Using ensembling to demystify fluctuations and correlations in GLMs trained on different, but correlated, features, in a new preprint by Bruno Loureiro, Cedric Gerbelot, Maria Refinetti, Gabriele Sicuro, and Krzakala Florent. On the ArXiv: arxiv.org/abs/2201.13383 #KCLDisSystResearch
Phase diagram of Stochastic Gradient Descent in high-dimensional two-layer neural networks deepai.org/publication/ph… by Rodrigo Veiga et al. including Bruno Loureiro #StochasticGradientDescent #ComputerScience
With @rodsveiga Bruno Loureiro Ludovic Stephan, and Krzakala Florent In the figure the axed are scaling exponents of the learning and of the network width with dimension.
You can learn convolutions from scratch, if you have the right implicit bias... in your training data! Very happy to share this collaboration with Alessandro Ingrosso, now on arXiv and soon at CosyneMeeting! Quick thread below 🧵👇
#Photodujour 📸 Portrait de Mrs. Frances Leyland (1871-1874) de James Abbott McNeill Whistler présentée dans l’accrochage « JamesMcNeill Whistler (1834 –1903). Chefs-d’œuvre de la Frick Collection, New York » au Musée d'Orsay 📷©The Frick Collection / Joseph Coscia Jr.
New preprint : arxiv.org/abs/2202.04509. When and how should you decay your learning rate ? We give some theoretical insights on this crucial question in our latest work with Maria Refinetti and Giulio Biroli. (1/3)
