Alexandre Adam @AlexandreAdam42

PhD Student, Physics, Université de Montréal Montréal Joined August 2022

Tweets

13
Followers

30
Following

276
Likes

369

Alexandre Adam @AlexandreAdam42

9 months ago

@deepcohen @alex_damian_ This is fascinating! Do you believe this theory can explain why taking the EMA of the weight while training is a useful thing to do? This is a strategy I know from the diffusion model litterature, unsure about wider applications. Cheers!

1 0 4 854 1

View Details

Jeremy Cohen @deepcohen

9 months ago

Even with full-batch gradients, DL optimizers defy classical optimization theory, as they operate at the *edge of stability.* With @alex_damian_, we introduce "central flows": a theoretical tool to analyze these dynamics that makes accurate quantitative predictions on real NNs.

18 211 1K 235K 1K

View Details

Alexandre Adam @AlexandreAdam42

9 months ago

@MiTiBennett The thesis is brillant! I can't believe how well written this is. A work of art

1 0 1 79 0

View Details

Alexandre Adam @AlexandreAdam42

9 months ago

@prsnltrdr @LucaAmb @martinmbauer Intuition

0 0 0 41 0

View Details

Alexandre Adam @AlexandreAdam42

10 months ago

@Kaju_Nut @martinmbauer There is a limit to our understanding. Solomonoff’s theory of inductive inference is what you are looking for. Finding the best theory is an intractable problem equivalent to the halting problem and Gödel’s incompleteness theorem.

1 1 1 217 0

View Details

Alexandre Adam @AlexandreAdam42

11 months ago

@LucaAmb @Kaju_Nut I think the problem is that Copenhagen’s interpretation is still widely taught and unnecessarily confusing. There should be an easier was to teach it that is more intuitive. Just like what path integral did for QFT.

0 0 0 28 0

View Details

Alexandre Adam @AlexandreAdam42

12 months ago

@memcculloch Denis Villeneuve, no questions.

0 0 0 28 0

View Details

Alexandre Adam @AlexandreAdam42

a year ago

@LucaAmb This is an awesome work!

1 0 1 536 0

View Details

Alexandre Adam @AlexandreAdam42

2 years ago

@Micah_mb010 Santa brought his gifts early this year

0 0 1 13 0

View Details

Ciela Institute @CielaInstitute

2 years ago

Score-based Likelihood Characterization (SLIC)-an ML method for tackling non-Gaussian noise in astronomical images! In our new blog post, Ronan Legin shows the ‘SLIC trick’ to learn the noise distribution from images enhancing parameter inference accuracy. ciela.science/2024/08/12/sco…

0 4 4 241 0

View Details

Alexandre Adam @AlexandreAdam42

2 years ago

@predict_addict This seems like a relevant topic. Bayesian inference does scale when you couple it to diffusion model, in which case it just becomes SOTA and solves problems impossible to even conceptualize from a frequentist perspective. Thanks for sharing, will check it out.

0 0 0 166 0

View Details

Ciela Institute @CielaInstitute

2 years ago

Exploring the cosmos using machine learning: Alexandre Adam from @ciela enhances Hubble images to JWST quality, revealing the untapped potential of images to convey vast information. Discover how modeling data as images revolutionizes cosmos understanding. ciela.science/2024/03/15/ech…