Research map / Probabilistic and scientific modeling / Molecules and proteins
Relative representations enable zero-shot latent space communication
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 |
|---|---|---|---|
| Represent each sample by cosine similarities to shared anchors instead of absolute coordinates | FastText and Word2Vec embeddings · English word embeddings, cross-model comparison | MRR: 0.00 ± 0.00 -> 0.94 ± 0.00 | Helped |
| Train with relative rather than absolute representations | Image and graph classification models · MNIST, Fashion-MNIST, CIFAR-10, CIFAR-100 coarse, Cora, Citeseer, PubMed | weighted F1: MNIST: 97.95 ± 0.10 -> 97.91 ± 0.07; CIFAR-100 coarse: 68.88 ± 0.14 -> 66.72 ± 0.35; Cora: 0.90 ± 0.01 -> 0.89 ± 0.02 | Mixed |
| Use relative representations when stitching independently trained autoencoder components | AE · MNIST reconstruction | MSE: 97.79 ± 2.48 -> 2.83 ± 0.20 | Helped |
| Use relative representations when stitching independently trained variational autoencoder components | VAE · MNIST reconstruction | MSE: 98.51 ± 1.49 -> 13.43 ± 6.79 | Helped |
| Use relative representations with translated parallel anchors | Cross-lingual RoBERTa stitching · Amazon coarse, English decoder and Spanish encoder | weighted F1: 43.67 ± 1.09 -> 82.78 ± 0.81 | Helped |
| Use relative representations when stitching across transformer architectures | BERT, ELECTRA, and RoBERTa · TREC coarse | weighted F1: 21.49 ± 3.64 -> 75.89 ± 5.38 | Helped |
| Use relative representations to stitch different image architectures | ViT-base-patch16-224 encoder and ViT-base-resnet50-384 decoder · CIFAR-100 coarse | weighted F1: 6.21 ± 0.33 -> 81.42 ± 0.38 | Helped |
| Increase the number of anchors (ablation) | Frozen image encoder using relative representations · CIFAR-100 coarse classification | accuracy | Helped |
| Increase the number of anchors (ablation) | End-to-end relative-representation model · Cora node classification | accuracy | Mixed |
| Select anchors using farthest point sampling rather than uniform sampling (ablation) | Relative word embeddings · FastText and Word2Vec cross-model comparison | Jaccard: 0.34 ± 0.01 -> 0.34 ± 0.01; 0.39 ± 0.00 -> 0.41 ± 0.00 | Helped |
| Select anchors using k-means rather than uniform sampling (ablation) | Relative word embeddings · FastText and Word2Vec cross-model comparison | Cosine: 0.86 ± 0.00 -> 0.87 ± 0.00 | Helped |
| Select anchors from the most frequent words, skipping the first 400 (ablation) | Relative word embeddings · FastText and Word2Vec cross-model comparison | MRR: 0.94 ± 0.00 -> 0.84 ± 0.01 | Hurt |
Closest papers by meaning
- Connecting Neural Models Latent Geometries with Relative Geodesic Representations (NeurIPS 2025 · similarity 0.91)
- From Bricks to Bridges: Product of Invariances to Enhance Latent Space Communication (ICLR 2024 · similarity 0.90)
- Latent Functional Maps: a spectral framework for representation alignment (NeurIPS 2024 · similarity 0.89)
- Metric Space Magnitude for Evaluating the Diversity of Latent Representations (NeurIPS 2024 · similarity 0.88)
- Joint‑Embedding vs Reconstruction: Provable Benefits of Latent Space Prediction for Self‑Supervised Learning (NeurIPS 2025 · similarity 0.87)
- Isometric Representation Learning for Disentangled Latent Space of Diffusion Models (ICML 2024 · 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.