Research map / Probabilistic and scientific modeling / Variational and generative sampling
Gradient Origin Networks
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 |
|---|---|---|---|
| replace the separate encoder with gradient-based origin encodings | GON · MNIST, Fashion-MNIST, Small NORB, CIFAR-10 | validation reconstruction loss (summed squared error): MNIST: 0.41\!\pm\!0.01 vs 1.33\!\pm\!0.02; Fashion-MNIST: 1.00\!\pm\!0.01 vs 1.75\!\pm\!0.02; Small NORB: 0.17\!\pm\!0.0 | Helped |
| replace the separate encoder with gradient-based origin encodings | GON · COIL20 | validation reconstruction loss (summed squared error): 5.35\!\pm\!0.01 vs AE 6.05\!\pm\!0.11; GLO 3.27\!\pm\!0.02 | Mixed |
| use a variational gradient-origin model instead of a VAE | VGON · MNIST, Small NORB, COIL20, CIFAR-10, CelebA | validation ELBO (bits/dim): MNIST: 1.06 vs 1.15; Small NORB: 2.34 vs 2.54; COIL20: 3.44 vs 3.63; CIFAR-10: 5.85 vs 5.94; CelebA: 5.41 vs 5.59 | Helped |
| use a variational gradient-origin model instead of a VAE | VGON · Fashion-MNIST | validation ELBO (bits/dim): 3.30 vs 3.23 | Hurt |
| detach gradients from the latent code before reconstruction (ablation) | GON · MNIST | validation reconstruction loss (summed squared error): 0.41\!\pm\!0.01 -> 8.17\!\pm\!0.15 | Hurt |
| detach gradients from the latent code and use 10 update steps (ablation) | GON · CIFAR-10 | validation reconstruction loss (summed squared error): 9.12\!\pm\!0.03 -> 28.68\!\pm\!1.16 | Hurt |
| use 10 latent update steps instead of one (ablation) | GON · CIFAR-10 | validation reconstruction loss (summed squared error): 9.12\!\pm\!0.03 -> 9.19\!\pm\!0.01 | No clear effect |
| use 10 latent update steps instead of one (ablation) | GON · MNIST | validation reconstruction loss (summed squared error): 0.41\!\pm\!0.01 -> 0.42\!\pm\!0.01 | No clear effect |
| initialize the latent input with random noise rather than the origin (ablation) | GON · CIFAR-10 | reconstruction loss | Hurt |
| use a gradient-descent-style latent update rather than the proposed negative-gradient encoding (ablation) | GON · CIFAR-10 | reconstruction loss | Hurt |
| use an activation whose second derivative is non-zero (ablation) | GON · image reconstruction | reconstruction loss | Helped |
| use ELU instead of ReLU (ablation) | GON · image reconstruction | reconstruction loss | Helped |
Closest papers by meaning
- Exploring the Latent Space of Autoencoders with Interventional Assays (NeurIPS 2022 · similarity 0.88)
- Perceptual Generative Autoencoders (ICML 2020 · similarity 0.87)
- From Variational to Deterministic Autoencoders (ICLR 2020 · similarity 0.87)
- DDMI: Domain-agnostic Latent Diffusion Models for Synthesizing High-Quality Implicit Neural Representations (ICLR 2024 · similarity 0.86)
- Hyper-Transforming Latent Diffusion Models (ICML 2025 · similarity 0.86)
- Latent Zoning Network: A Unified Principle for Generative Modeling, Representation Learning, and Classification (NeurIPS 2025 · similarity 0.86)
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.