#iclr2027
The two (out of five) reviewers who gave me 2 haven't responded to my rebuttal. Kindly remind them, or let it slide so it won't get worse? (I'm with 3,3,3,2,2)
I have trauma from an ICML review that, after the rebuttal, went from 2 to 1 (paper sitll got in).
What a week at #ICML2026!
I had the opportunity to present two papers:
โข SONAR (ICML Main Track), where we study how spectral bias affects both audio generators and detectors and introduce a contentโnoise representation learning framework for generalizable audio deepfake detection.
โข Catastrophic Forgetting is Low-Rank at the Continual Adaptation at Scale Workshop, where we develop a function-space view of continual learning, showing that forgetting can be predicted through NTK interactions and often concentrates in a surprisingly small subspace.
Grateful for all the conversations, feedback, and people I met along the way. These discussions were one of the highlights of the conference!
If you're interested in chatting about model inductive biases, NTKs, continual learning and forgetting, or deepfake detection, feel free to reach out.
#DeepLearning #ICML2026 #ContinualLearning #DeepfakeDetection #NTK