Research map / Probabilistic and scientific modeling / Variational and generative sampling
Exponential Family Embeddings
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 |
|---|---|---|---|
| Model each neuron from the activities of nearby neurons in its context | Gaussian embeddings · Zebrafish neural activity, single neuron and 25% held out | mean squared error: FA: 0.290±0.003; Gaussian embedding (c=50, K=100): 0.222±0.002 (single neuron); 0.232 ± 0.003 (25% held out) | Helped |
| Increase context size from c=10 to c=50 (ablation) | Gaussian embeddings · Zebrafish neural activity | mean squared error: 0.239±0.006 -> 0.227±0.002 (K=10, single neuron); 0.246 ± 0.004 -> 0.235 ± 0.003 (K=10, 25% held out) | Helped |
| Increase latent dimension from K=10 to K=100 (ablation) | Gaussian embeddings with c=50 · Zebrafish neural activity | mean squared error: 0.227±0.002 -> 0.222±0.002 (single neuron); 0.235 ± 0.003 -> 0.232 ± 0.003 (25% held out) | Helped |
| Constrain the Gaussian embedding vectors to be nonnegative (ablation) | Gaussian embeddings · Zebrafish neural activity, single neuron held out | mean squared error: 0.239±0.006 -> 0.263±0.004 (c=10, K=10); 0.222±0.002 -> 0.261±0.004 (K=100 comparison) | Hurt |
| Downweight zero entries by a factor of 0.1 (ablation) | Poisson embeddings · Market basket held-out data | normalized predictive log-likelihood: -7.497±0.007 -> -7.110±0.007 (K=20); -7.199±0.008 -> -6.950±0.007 (K=100) | Helped |
| Downweight zero entries by a factor of 0.1 (ablation) | Poisson embeddings · MovieLens-100K held-out ratings | normalized predictive log-likelihood: -5.691±0.006 -> -5.790±0.003 (K=20); -5.726±0.005 -> -5.798±0.003 (K=100) | Hurt |
| Use Poisson embeddings rather than HPF or Poisson PCA | Poisson embeddings · Market basket held-out data | normalized predictive log-likelihood: -7.199±0.008 vs. HPF -7.626±0.007 and Poisson PCA -11.01±0.01 (K=100) | Helped |
| Use Poisson embeddings rather than HPF or Poisson PCA | Poisson embeddings · MovieLens-100K held-out ratings | normalized predictive log-likelihood: -5.691±0.006 vs. HPF -5.787±0.006 and Poisson PCA -5.908±0.006 (K=20) | Helped |
| Replace the multiplicative Poisson parameterization with the additive parameterization (ablation) | Poisson embeddings · Market basket held-out data | normalized predictive log-likelihood: -7.497±0.007 -> -7.868±0.005 (K=20); -7.199±0.008 -> -8.414±0.003 (K=100) | Hurt |
| Replace the multiplicative Poisson parameterization with the additive parameterization (ablation) | Poisson embeddings · MovieLens-100K held-out ratings | normalized predictive log-likelihood: -5.691±0.006 -> -5.964±0.003 (K=20); -5.726±0.005 -> -6.118±0.002 (K=100) | Hurt |
| Increase latent dimension from K=20 to K=100 (ablation) | Poisson embeddings · MovieLens-100K held-out ratings | normalized predictive log-likelihood: -5.691±0.006 -> -5.726±0.005 | Hurt |
Closest papers by meaning
- Structured Embedding Models for Grouped Data (NeurIPS 2017 · similarity 0.91)
- Zero-Inflated Exponential Family Embeddings (ICML 2017 · similarity 0.89)
- Context Selection for Embedding Models (NeurIPS 2017 · similarity 0.88)
- By Tying Embeddings You Are Assuming the Distributional Hypothesis (ICML 2024 · similarity 0.86)
- Demystifying Embedding Spaces using Large Language Models (ICLR 2024 · similarity 0.85)
- When is an Embedding Model More Promising than Another? (NeurIPS 2024 · 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.