Steffen Rulands
@rulands.bsky.social
Professor at LMU Munich. Theoretical physics, biophysics, artificial intelligence
We derive a theory from first principles that predicts that the homogeneous state of deep neural networks is unstable in a way that leads to the emergence of periodic channel structures.
January 14, 2025 at 11:10 AM
We derive a theory from first principles that predicts that the homogeneous state of deep neural networks is unstable in a way that leads to the emergence of periodic channel structures.
New preprint: are deep neural networks just complicated curve fitters or can they show emergent behavior that is independent of the data? We show that training deep neural networks gives rise to emergent weight morphologies independent of the training data.
arxiv.org/abs/2501.05550
arxiv.org/abs/2501.05550
January 14, 2025 at 11:10 AM
New preprint: are deep neural networks just complicated curve fitters or can they show emergent behavior that is independent of the data? We show that training deep neural networks gives rise to emergent weight morphologies independent of the training data.
arxiv.org/abs/2501.05550
arxiv.org/abs/2501.05550
New preprint: Ageing is tightly reflected in longitudinal evolution of DNA methylation marks. We show that ageing also leaves a stereotypic footprint in how DNA methylation is correlated in the genome: positions that are roughly 500bp apart become more correlated. www.biorxiv.org/content/10.1...
December 18, 2024 at 2:57 PM
New preprint: Ageing is tightly reflected in longitudinal evolution of DNA methylation marks. We show that ageing also leaves a stereotypic footprint in how DNA methylation is correlated in the genome: positions that are roughly 500bp apart become more correlated. www.biorxiv.org/content/10.1...
Unfortunately, today in the lecture Boltz-man was brutally beaten up by S-man.
November 29, 2024 at 3:57 PM
Unfortunately, today in the lecture Boltz-man was brutally beaten up by S-man.
Teaching statistical physics this term. If things get difficult, we get help from super heroes like Mr T (who can relate macroscopic variables) or S-man (who can derive them micrscopically). There is also a villain called Boltz-man who has the habit of creating super-large numbers.
November 16, 2024 at 2:46 PM
Teaching statistical physics this term. If things get difficult, we get help from super heroes like Mr T (who can relate macroscopic variables) or S-man (who can derive them micrscopically). There is also a villain called Boltz-man who has the habit of creating super-large numbers.