Show HN: I invented a new generative model and got accepted to ICLR
11 hours ago
- #ICLR 2025
- #generative models
- #machine learning
- Discrete Distribution Networks (DDN) introduced as a novel generative model with hierarchical discrete distributions.
- DDN fits target distributions by generating multiple discrete sample points, refining outputs layer by layer.
- Unique properties of DDN include general zero-shot conditional generation and 1D latent representation.
- Experiments demonstrate DDN's efficacy on datasets like CIFAR-10 and FFHQ.
- DDN supports zero-shot conditional generation across non-pixel domains without relying on gradients.
- Training phase shows generated images become increasingly similar to training images as network depth increases.
- Split-and-Prune strategy reduces KL divergence more effectively than Gradient Descent alone.
- Future research directions include hyperparameter tuning, scaling to ImageNet, and applying DDN to language modeling.
- DDN's GPU memory requirements are slightly higher than conventional GANs but manageable.
- DDN avoids mode collapse by selecting outputs most similar to GT and using L2 loss.