Alberto Bernacchia (@AlbertoBernacc1) profile picture

Neural networks | University of Cambridge | now at MediaTek Research

Joined May 2019
Photo by Alberto Bernacchia (@AlbertoBernacc1)Photo by Alberto Bernacchia (@AlbertoBernacc1)Photo by Alberto Bernacchia (@AlbertoBernacc1)Photo by Alberto Bernacchia (@AlbertoBernacc1)Photo by Alberto Bernacchia (@AlbertoBernacc1)Photo by Alberto Bernacchia (@AlbertoBernacc1)
This is NeurIPS week!!! @MediaTek Research has 4 papers accepted + 2 at workshops (links below) Please note that *WE ARE HIRING*, please ping me if you are interested. Continual Learning Meta-Learning Gaussian Processes Bayesian Optimization Computer Vision Language Models
1
5
ICLR reviews just came out. You can find here all 3264 ICLR paper submissions ranked by average score: dropbox.com/s/kbue3dz5ch6k03…
1
Kafkaesque tale Last night I was supposed to catch a Ryanair flight from Frankfurt to London. At the time of boarding they do not let me in: my antigen test is three days old, while the limit is two days, they say.
2
4
The cops finally conclude that I'm right, they call Ryanair and say that I should be allowed to board. Ryanair replies that it doesn't care, Ryanair decides who boards and who doesn't. The plane takes off. The cops tell me that I can sleep on the airport chairs if I want.
1
In the meantime, Germany today has 400 Covid cases, England has 24,000, and I had the second dose of vaccine two weeks ago. At Wembley 40,000 fans are screaming face to face, and people go to restaurants without a mask as if Covid no longer exists.
"Collusion rings threaten the integrity of computer science research." This is so obvious, I wrote a detailed feedback to ICML organisers about this. I got to review precisely the papers I wanted. I don't understand how they spotted this just now. dl.acm.org/doi/10.1145/34297…
1
1
Finally I have my own "Schmidhuber" moment. This ICLR 2021 oral has the same title of a paper that I published 9 years ago!!! (They cite my paper, so it's OK.) openreview.net/forum?id=RGJb…
7
Have you ever considered training your model with a negative learning rate? My #ICLR2021 paper shows that you should (if your model is meta-learning). openreview.net/forum?id=60j5… 1/5 👇
2
1
14
I used MAML, which finds the optimal initial weights for adapting to multiple tasks. My result suggests that the optimal initial weights are found by pushing away from the optimum, rather than towards, for each task. 4/5
1
Thanks @ritwikniyogi and Paolo Grazieschi for helping with the revision!
2