Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

The possibility of using maximum likelihood estimation for training #26

Open
daihuiao opened this issue Jun 5, 2024 · 1 comment
Open

Comments

@daihuiao
Copy link

daihuiao commented Jun 5, 2024

Hello,

We understand that using computed likelihood directly for maximum likelihood estimation can be computationally intensive. However, we are curious if this technique is theoretically and practically feasible. We would like to experiment with this approach on smaller models and datasets. Any advice or code assistance would be greatly appreciated.
:)
Once again, thank you for your outstanding work.

@daihuiao
Copy link
Author

I use the code in the repository to calculate likelihood for maximum likelihood training. As a result, the pre-training of score function is lost, and the effect of training directly with likelihood is poor.
What you normally learn with the socre function is an irregular distribution (that's what data sets look like).

image

But what is learned by calculating the likelihood is its distribution after collapsing towards the middle:

image
Like Gaussian distributions or VAE, learned strategies can only be learned up to a peak.

Anyone with other ideas for improvement?

@daihuiao daihuiao reopened this Jul 24, 2024
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
None yet
Projects
None yet
Development

No branches or pull requests

1 participant