Research map / Architectures and efficiency / RNNs and normalization
Universal Hopfield Networks: A General Framework for Single-Shot Associative Memory Models
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 Manhattan distance similarity instead of dot-product similarity (ablation) | UHN with softmax separation · MNIST | retrieval capacity | No clear effect |
| use Manhattan distance similarity instead of dot-product similarity (ablation) | UHN with softmax separation · CIFAR10 | retrieval capacity | Helped |
| use Manhattan distance similarity instead of dot-product similarity (ablation) | UHN with softmax separation · Tiny ImageNet | retrieval capacity | Helped |
| use Euclidean distance similarity instead of dot-product similarity (ablation) | UHN with softmax separation · image datasets | retrieval capacity | Helped |
| use KL, Jensen-Shannon, or reverse KL similarity instead of Euclidean, dot-product, or Manhattan similarity (ablation) | UHN with softmax separation · image datasets | retrieval capacity | Hurt |
| use dot-product similarity instead of Manhattan distance (ablation) | UHN with softmax separation · Gaussian-noise-corrupted images | retrieval robustness | Helped |
| use Manhattan distance similarity instead of dot-product similarity (ablation) | UHN with softmax separation · masked images | retrieval robustness | Mixed |
| use dot-product similarity instead of Manhattan distance (ablation) | UHN with softmax separation · Tiny ImageNet with extremely high masking | retrieval robustness | Helped |
| use a high-order polynomial separation function instead of softmax with beta set to 1 (ablation) | UHN with dot-product similarity · image retrieval capacity | retrieval capacity | Helped |
| use exponential or max separation instead of lower-order separation functions (ablation) | UHN with dot-product similarity · image retrieval capacity | retrieval capacity | Helped |
| increase softmax beta toward the max-separation regime (ablation) | UHN with dot-product similarity · image retrieval capacity | retrieval capacity | Helped |
| replace dot-product attention similarity with Manhattan or Euclidean similarity (ablation) | transformer · WikiText | perplexity | No clear effect |
Closest papers by meaning
- Uniform Memory Retrieval with Larger Capacity for Modern Hopfield Models (ICML 2024 · similarity 0.90)
- Dense Associative Memory Through the Lens of Random Features (NeurIPS 2024 · similarity 0.89)
- Relating Hopfield Networks to Episodic Control (NeurIPS 2024 · similarity 0.89)
- Sparse and Structured Hopfield Networks (ICML 2024 · similarity 0.89)
- Nonparametric Modern Hopfield Models (ICML 2025 · similarity 0.88)
- On Computational Limits of Modern Hopfield Models: A Fine-Grained Complexity Analysis (ICML 2024 · similarity 0.88)
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.