Research map / Probabilistic and scientific modeling / Variational and generative sampling
Reliable Measures of Spread in High Dimensional Latent Spaces
Anna Marbut, Katy McKinney-Bock, Travis J Wheeler
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 |
|---|---|---|---|
| Use EEE instead of ACS or I(V) to compare spread | Synthetic latent-space distributions · Seven synthetic distributions across dimensions | agreement with expected relative spread ordering | Improved |
| Use VRM instead of ACS or I(V) to compare spread | Synthetic latent-space distributions · Seven synthetic distributions across dimensions | agreement with expected relative spread ordering | Improved |
| Use EEE rather than Eigenvalue Ratio (ER) | Principal-component spread measurement · 100D cluster distributions | relative spread discrimination: 0.5394 vs 0.0001; 0.6746 vs 0.0007 | Improved |
| Use average cosine similarity to assess spread | Latent-space distributions · Symmetric Clusters vs Sphere | relative spread distinction: 0.0007 vs 0.0007 | Hurt |
| Use I(V) to distinguish Cone from other spherical distributions | Latent-space distributions · 100D Cone vs Sphere | I(V): 0.9944 vs 0.9988 | Hurt |
| Add random uniform noise to Word2Vec embeddings | Pre-trained Word2Vec embeddings · Word2Vec Google News | EEE: 0.4058 -> 0.0367 | Improved |
| Add random uniform noise to Word2Vec embeddings | Pre-trained Word2Vec embeddings · Word2Vec Google News | VRM: 0.1525 -> 0.0010 | Improved |
| Add random uniform noise to Word2Vec embeddings | Pre-trained Word2Vec embeddings · Word2Vec Google News | ACS: 0.1317 -> 0.0022 | Improved |
| Add random uniform noise to Word2Vec embeddings | Pre-trained Word2Vec embeddings · Word2Vec Google News | I(V): 0.9251 -> 0.9996 | Improved |
| Use nearest-neighbor entropy measures based wholly on Euclidean distances | NNR and NNKL · Sphere, Cone, and Nested Shell, especially at high dimensions | relative spread reliability: 100D NNR: 1.0023 vs 1.0022; 100D NNKL: 0.2558 vs 0.2853 | Hurt |
| Use KNN Overlap as a relative spread measure | Synthetic latent-space distributions · High-dimensional synthetic distributions | relative spread discrimination: 100D: 0.0078 vs 0.8491 | Hurt |
| Use KL-divergence-based measures for comparisons across dimensionalities | DKLM and NNKL · Synthetic distributions across dimensions | cross-dimensional comparability: 100D NNKL cluster scores are almost 300x the 2-dimensional counterparts | Hurt |
Closest papers by meaning
- Metric Space Magnitude for Evaluating the Diversity of Latent Representations (NeurIPS 2024 · similarity 0.85)
- Diff-eRank: A Novel Rank-Based Metric for Evaluating Large Language Models (NeurIPS 2024 · similarity 0.84)
- On the Entropy Calibration of Language Models (NeurIPS 2025 · similarity 0.84)
- Correlation Dimension of Autoregressive Large Language Models (NeurIPS 2025 · similarity 0.84)
- Continuous Diffusion Model for Language Modeling (NeurIPS 2025 · similarity 0.83)
- Probing the Latent Hierarchical Structure of Data via Diffusion Models (ICLR 2025 · similarity 0.83)
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.
