Madhu Advani (@advani_madhu) 's Twitter Profile
Madhu Advani

@advani_madhu

ML researcher/engineer @Apple, Postdoc @Harvard, PhD in Applied physics @Stanford

ID: 1074670632

calendar_today09-01-2013 19:33:32

6 Tweet

208 Followers

63 Following

Madhu Advani (@advani_madhu) 's Twitter Profile Photo
Madhu Advani
@advani_madhu

6 months ago

Our new theory paper on self-supervised learning is on arxiv! We show how JEPA and self-distillation networks, which do reconstruction in the latent space, are implicitly biased to avoid noisy feature directions in a way that masked auto-encoders (MAEs) are not.

Our new theory paper on self-supervised learning is on arxiv! We show how JEPA and self-distillation networks, which do reconstruction in the latent space, are implicitly biased to avoid noisy feature directions in a way that masked auto-encoders (MAEs) are not.