Guys, how can use your method let's say for NN model?

[diff7]

Hi,
I was looking at your method and let's say I want to apply for NN. But I have not figured out how to construct matrix Phi ( A in the code ).

Any tips and directions will be appreciated!

[jackyzyb]

Hi,
Thanks for asking! Our method can be applied to Bayesian NNs. Here's a good reference to construct Phi and Bayesian coresets on Bayesian NNs: this work (https://github.com/rpinsler/active-bayesian-coresets) uses a resnet with Bayesian final layer, and the matrix Phi can be sampled from the posterior. They use the FW method (https://papers.nips.cc/paper/2019/file/84c2d4860a0fc27bcf854c444fb8b400-Paper.pdf) to construct Bayesian coresets, which also needs to construct the same matrix Phi.

[diff7]

Thank you!

If you do not mind I will keep the issue open, I am slowly digging into it and may have more questions ))

By the way, do you refer to equation 16 in the paper?

[jackyzyb]

Hi! Sorry I must be too busy around Feb 1 so that I missed all your comments...
Yes it is. The \hat{L}_n here corresponds to the n^th column in the matrix Phi.
Feel free to ask more questions!

[diff7]

Got it, thanks!