Thanks to your repo and pseudo code in the manuscript, I could implement most experiments except for the local energy.
My question is, is the decomposing kernel at time t or at time 0? For the definition of NTK, it should be at time 0, but the pre-trained model may not have much energy concentration per your work 'What can linearized neural networks actually say about generalization?'. I also do eigenfunction decomposition as your repo did before, would you mind sharing more details with me?
It'll be very kind of you if you'd like to discuss this with me, and I'll be more than grateful if you'd like to share some code. Thanks in advance!
Thanks to your repo and pseudo code in the manuscript, I could implement most experiments except for the local energy.
My question is, is the decomposing kernel at time t or at time 0? For the definition of NTK, it should be at time 0, but the pre-trained model may not have much energy concentration per your work 'What can linearized neural networks actually say about generalization?'. I also do eigenfunction decomposition as your repo did before, would you mind sharing more details with me?
It'll be very kind of you if you'd like to discuss this with me, and I'll be more than grateful if you'd like to share some code. Thanks in advance!