About RD model

[Lupin123123]

Hi!

I recently read your paper, "An Improved Upper Bound on the Rate-Distortion Function of Images", and found it truly inspiring. I do, however, have a few questions regarding the model settings discussed in Table 3. Specifically, if the model size continues to increase, will the rate-distortion performance eventually converge, given that the RD function itself is independent of the model?

[duanzhiihao]

Hi, thank you for your interest in our work!

Specifically, if the model size continues to increase, will the rate-distortion performance eventually converge, given that the RD function itself is independent of the model?

Yes, it will converge to the information RD function assuming several conditions (e.g., a super-large model can approximate any function, the dataset used for training/evaluation is large enough to represent the image distribution, etc).

[Lupin123123]

Hi,
I was also wondering whether the estimated rate-distortion function (or its upper bound) obtained by a VAE is achievable for an arbitrary source?

[duanzhiihao]

Yes if the VAE is trained and tested on that source.