Research map / Label-efficient and robust learning / Recognition and long-tail
Geometrically Aligned Transfer Encoder for Inductive Transfer in Regression Tasks
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes GATE, a differential-geometry-inspired transfer encoder that aligns task-specific latent coordinate patches via transfer maps, consistency, mapping, and perturbation-distance losses for inductive transfer in molecular property regression.
Problem
Inductive transfer learning for regression tasks, where most transfer methods target classification on images and text; the paper focuses on small molecular property datasets.
Approach
Interpret encoder latent spaces as Riemannian manifolds, learn diffeomorphism-like transfer maps between source and target task coordinate patches with autoencoder and consistency losses, and regularize with a distance loss on infinitesimal perturbation points so that overlapping regions are locally flat and aligned.
Key results
- GATE achieved the best RMSE in 12 of 23 task pairs under random split and 13 of 23 under scaffold split, versus at most 5 pairs for any other method.
- Average RMSE was 9.8% lower than the second-best method (GSP-KD) under random split and 14.3% lower than MTL under scaffold split.
- Adding the distance loss reduced overfitting and improved robustness to corrupted training labels compared with MTL.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use GATE instead of GSP-KD | molecular property regression transfer · 23 task pairs, random split | average RMSE: 9.8 \% lower | Helped |
| use GATE instead of MTL | molecular property regression transfer · 23 task pairs, scaffold split | average RMSE: 14.3 \% lower | Helped |
| add distance loss to mapping and consistency losses (ablation) | GATE · molecular property regression | overfitting during training | Helped |
| add distance loss to mapping and consistency losses (ablation) | GATE · molecular property regression | regression task performance | No clear effect |
| use GATE instead of MTL | regression with four different source tasks · latent-space stability | latent-space stability | Helped |
| use GATE instead of MTL | molecular property regression under data corruption · corrupted training data points | MSE against original values | Helped |
Surprising findings
- Knowledge distillation and GSP-KD sometimes beat GATE on individual task pairs but show critically low accuracy on others.
Limitations
- Evaluated only on molecular property regression; generality to other domains is untested.
- Uses only infinitesimal perturbations and a flat-space distance approximation rather than solving geodesic equations.
- The Riemannian framing is heuristic; the paper does not formally verify the manifold assumptions.
Experimental setup
Datasets: PubChem property datasets OChem property datasets CCCB polarizability dataset
Benchmarks: 14 molecular property datasets (AS, BP, CCS, CT, DK, DS, EF, IP, KRI, LP, POL, ST, VS, HV)
Metrics: RMSE relative RMSE
Baselines: single task learning multi-task learning knowledge distillation GSP-KD transfer learning (retrain all) transfer learning (head only)
Builds on: DMPNN GSP-KD
Closest papers by meaning
- Graph Geometry-Preserving Autoencoders (ICML 2024 · similarity 0.85)
- Function Encoders: A Principled Approach to Transfer Learning in Hilbert Spaces (ICML 2025 · similarity 0.85)
- Disentangled Representation Learning with the Gromov-Monge Gap (ICLR 2025 · similarity 0.84)
- Toward a Unified Geometry Understanding : Riemannian Diffusion Framework for Graph Generation and Prediction (NeurIPS 2025 · similarity 0.83)
- Geometry Awakening: Cross-Geometry Learning Exhibits Superiority over Individual Structures (NeurIPS 2024 · similarity 0.83)
- DisRFM: Polar Riemannian Flow Matching for Structure-Preserving Graph Domain Adaptation (NeurIPS 2026 · 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.