Research map / Probabilistic and scientific modeling / Variational and generative sampling
Evidential Softmax for Sparse Multimodal Distributions in Deep Generative Models
Phil Chen, Masha Itkina, Ransalu Senanayake, Mykel Kochenderfer
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Train with sparse, multimodal Ev-softmax rather than Entmax-1.5 | Semi-supervised VAE · MNIST | accuracy (%): 97.20 ± .001 -> 97.30 ± .001 | Improved |
| Use Ev-softmax rather than Sparsemax | Semi-supervised VAE · MNIST | decoder calls: 1.01 ± 0.01 -> 1.64 ± 0.01 | Hurt |
| Train the PixelCNN prior with Ev-softmax rather than Sparsemax | VQ-VAE · tinyImageNet | top-5 accuracy: 40.0 ± 0.2 -> 43.6 ± 0.3 | Improved |
| Train the PixelCNN prior with Ev-softmax rather than Sparsemax | VQ-VAE · tinyImageNet | top-10 accuracy: 52.0 ± 0.2 -> 55.6 ± 0.2 | Improved |
| Use Ev-softmax rather than Sparsemax | VQ-VAE · tinyImageNet | average number of nonzero latent classes: 46 ± 0.3 -> 77 ± 0.2 | Mixed |
| Use Ev-softmax rather than Softmax | Transformer self-attention · IWSLT 2014 EN→DE | BLEU: 29.2 ± 0.06 -> 29.4 ± 0.05 | Improved |
| Use sparse Ev-softmax attention rather than attention over the full source | Transformer self-attention · IWSLT 2014 EN→DE | attended source words: 39.5 ± 11.5 -> 8.2 ± 1.3 | Mixed |
| Train the CVAE with Ev-softmax rather than Softmax | CVAE · MNIST even/odd generation | Wasserstein distance | Improved |
| Train the CVAE with Ev-softmax rather than Sparsemax or Entmax-1.5 | CVAE · MNIST even/odd generation | Wasserstein distance | Improved |
| Train with sparse and multimodal outputs rather than applying evidential sparsification only after training | Semi-supervised VAE · MNIST | accuracy (%): 96.90 ± .001 -> 97.30 ± .001 | Improved |
| Train the prior with sparse normalization rather than Softmax | VQ-VAE PixelCNN prior · tinyImageNet | top-5 accuracy: 38.4 ± 0.2 -> 43.6 ± 0.3 | Improved |
| Retain more attended source words with Ev-softmax rather than Sparsemax | Transformer self-attention · IWSLT 2014 EN→DE | BLEU: 29.0 ± 0.05 -> 29.4 ± 0.05 | Improved |
Closest papers by meaning
- A Likelihood Based Approach to Distribution Regression Using Conditional Deep Generative Models (ICML 2025 · similarity 0.87)
- MultiMax: Sparse and Multi-Modal Attention Learning (ICML 2024 · similarity 0.86)
- Generative Marginalization Models (ICML 2024 · similarity 0.86)
- SoftCVI: Contrastive variational inference with self-generated soft labels (ICLR 2025 · similarity 0.85)
- Aggregation of Dependent Expert Distributions in Multimodal Variational Autoencoders (ICML 2025 · similarity 0.85)
- SoftVQ-VAE: Efficient 1-Dimensional Continuous Tokenizer (CVPR 2025 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
