We proposed to use Determinantal Point Process as a diversity encouraging prior for latent variable models, here Variational Auto-encoderto in particular, to alleviate imbalance learning problem.
The training examples include 5000 MNIST ’0’, ’1’ handwritten digits data, where digit ‘1’ is the minor class. In this demo, we use an imbalance ratio 10 to 1.
Run unbalance_vae_generator.py to generate synthetic hand-written digits using standard VAE.
Run unbalance_dppvae_generator.py to generate synthetic hand-written digits using the proposed VAE with Determinantal Point Process as the prior.
A comparison of results are shown below, where the proposed DPP-VAE generated more minor class '1':
| Synthetic data with standard VAE | Synthetic data with DPP VAE |
|---|---|
 |
 |