Niccolò Ajroldi
@n_ajroldi
Followers
163
Following
1K
Statuses
141
Research Engineer at Ellis Institute Tübingen and Max Planck Institute for Intelligent Systems
Milano, Lombardia
Joined May 2013
RT @ShashwatGoel7: 🚨Great Models Think Alike and this Undermines AI Oversight🚨 New paper quantifies LM similarity (1) LLM-as-a-judge favor…
0
25
0
RT @mvpatel2000: EMA in CPU is a really neat trick. People do similar async offloading for efficient checkpointing, but I've never seen EMA…
0
1
0
No one can save Elon from himself anymore
Only the AfD can save Germany
0
0
2
RT @orvieto_antonio: Hey you, fantastic machine learning researcher; what about a last-minute application for our PI call at our ELLIS Inst…
0
5
0
RT @ELLISInst_Tue: 🎙The fifth episode of the @Cyber_Valley Podcast featuring our Principal Investigators is now live! 🎉 @FrankRHutter dives…
0
1
0
RT @frankstefansch1: On Feb 11 & 12, Meta will host the first AlgoPerf Workshop on "How can we train neural nets faster?" featuring talks b…
0
10
0
RT @docmilanfar: As our fish-like ancestors evolved, developing necks, the nerve got stuck in its pathway. Evolution couldn't re-route the…
0
135
0
RT @ELLISInst_Tue: The new call for Principal Investigators at the ELLIS Institute Tübingen is out! 🚀 We are recruiting Principal Investig…
0
9
0
Great work by @ruuustem_10 that I had the pleasure to contribute to! We introduce a new function class that better captures neural network loss landscapes, ensuring convergence for several SGD-based algorithms, and showing its applicability across many Deep Learning tasks!
I am happy to announce that our recent publication "Loss Landscape Characterization of Neural Networks without Over-Parametrziation " that has been recently accepted to NeurIPS 2024 is now available on arXiv: 1/n
0
0
2
@jeankaddour @MLCommons Playing around with LAWA was fun! We got promising performances, but the current state of the competition API made it cumbersome to aggregate models and convert the speedup to a good final score. Hopefully future iterations of the benchmark would make LAWA shine more.
0
0
2
@orvieto_antonio @jonasgeiping @ELLISInst_Tue @aaron_defazio Our methodology is not novel, neither particularly sophisticated, but we hope it provides interesting insights into LR schedules and help develop the future iterations of AlgoPerf. Big thanks to @MLCommons for organizing this, and to our team at @ELLISInst_Tue! n/n
0
0
0
RT @frankstefansch1: The inaugural AlgoPerf results are in, highlighting a new generation of neural net training algorithms! Get 28% faster…
0
9
0