VAE: The Evidence Lower Bound
Intermediate
The problem, from first principles
The VAE maximizes a lower bound on the log-likelihood of the data. The bound is the reconstruction term minus a KL penalty, and it can be estimated with samples from the posterior.
From theory to code
Implement elbo(log_px_given_z, kl), returning the reconstruction term minus the KL.
Constraints
- The result is maximized; training minimizes its negative.
Hints
Hint 1
Subtract the KL from the reconstruction log-likelihood.
Keywords
research-papersunsupervisedvaelatent-variables
Python
Run your code to see results
No results yet
Run your code to check it against the examples, or submit to run every test.