Research map / Probabilistic and scientific modeling / Variational and generative sampling
Implicit MLE: Backpropagating Through Discrete Exponential Family Distributions
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 |
|---|---|---|---|
| Use I-MLE with Sum-of-Gamma perturbations and tuned temperature instead of MAP-only marginal approximation (ablation) | I-MLE for BeerAdvocate Aroma · BeerAdvocate Aroma, k=10 | Subset precision: 14.55 -> 47.98 | Helped |
| Replace local Gumbel perturbations with Sum-of-Gamma perturbations (ablation) | I-MLE for BeerAdvocate Aroma · BeerAdvocate Aroma, k=10 | Subset precision: 39.28 -> 47.98 | Helped |
| Tune the perturbation temperature rather than use MAP-only approximation (ablation) | I-MLE for BeerAdvocate Aroma · BeerAdvocate Aroma, k=10 | Test MSE: 4.08 -> 2.71 | Helped |
| Use I-MLE (μ-μ) instead of BB | Warcraft shortest path model · Warcraft shortest path, K=12 | Accuracy: 95.2\std{0.7} -> 97.2\std{0.5} | Helped |
| Use I-MLE (μ-μ) instead of BB | Warcraft shortest path model · Warcraft shortest path, K=18 | Accuracy: 94.7\std{0.4} -> 95.8\std{0.7} | Helped |
| Use I-MLE (μ-μ) instead of BB | Warcraft shortest path model · Warcraft shortest path, K=24 | Accuracy: 93.8\std{0.3} -> 94.3\std{1.0} | Helped |
| Use I-MLE (μ-μ) instead of BB | Warcraft shortest path model · Warcraft shortest path, K=30 | Accuracy: 93.6\std{0.5} -> 93.6\std{0.4} | No clear effect |
| Use I-MLE (μ-μ) instead of I-MLE (M-M) (ablation) | I-MLE for Warcraft shortest path · Warcraft shortest path | Accuracy: 95.2\std{0.3} -> 97.2\std{0.5}; 94.4\std{0.5} -> 95.8\std{0.7}; 93.2\std{0.2} -> 94.3\std{1.0}; 93.7\std{0.6} -> 93.6\st | Helped |
| Use Sum-of-Gamma instead of standard Gumbel perturbations (ablation) | I-MLE synthetic optimization · Synthetic 5-subset optimization | Final loss | Helped |
| Use I-MLE instead of the straight-through estimator | Discrete 10-subset VAE · MNIST discrete 10-subset VAE | Test loss | Helped |
| Use Sum-of-Gamma perturbations instead of standard Gumbel perturbations (ablation) | I-MLE discrete 10-subset VAE · MNIST discrete 10-subset VAE | Test loss | Helped |
Closest papers by meaning
- Adaptive Perturbation-Based Gradient Estimation for Discrete Latent Variable Models (AAAI 2023 · similarity 0.87)
- Exponential Family Estimation via Adversarial Dynamics Embedding (NeurIPS 2019 · similarity 0.86)
- Scalable Discrete Diffusion Samplers: Combinatorial Optimization and Statistical Physics (ICLR 2025 · similarity 0.86)
- DNA-SE: Towards Deep Neural-Nets Assisted Semiparametric Estimation (ICML 2024 · similarity 0.86)
- A Diffusion Model Framework for Unsupervised Neural Combinatorial Optimization (ICML 2024 · similarity 0.85)
- Joint Learning of Energy-based Models and their Partition Function (ICML 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.