Research map / Probabilistic and scientific modeling / Variational and generative sampling
Active Learning of Conditional Generative Models via the Transport Neural Tangent Kernel
Jayoung Ryu, Kyunghyun Cho, Romain Lopez
Read the paperCheck your idea against 137,217 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 an active learning method for conditional generative models of distributional experimental outputs, using a transport neural tangent kernel to guide experiment selection.
Problem
Automated experimental design for high-throughput experiments with distributional outputs, where acquiring data at new conditions is expensive and existing active learning frameworks make restrictive assumptions about the output distribution.
Approach
Introduces a transport neural tangent kernel to active learning of conditional generative models, enabling selection of new experimental conditions for distributional outputs rather than assuming simpler output models.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use tNTK instead of mean-based NTK for acquisition | GP total-variance active learning · Gaussian mixture (quantile, 5+1) | change in Wasserstein test risk (x1e-3): -416.43 -> -429.52 | Improved |
| use tNTK instead of mean-based NTK for acquisition | GP total-variance active learning · Dew Point (quantile, 5+1) | change in Wasserstein test risk (x1e-3): -178.44 -> -221.18 | Improved |
| use tNTK instead of NTK | conditional flow matching AL · Radial flow, CLINC150, Van der Pol | change in Sinkhorn divergence: CLINC150 5+1: -306.01 -> -672.85 | Improved |
| use tNTK instead of NTK at a larger budget | conditional flow matching AL · CLINC150 (5+5) | change in Sinkhorn divergence: -884.02 vs -852.31 | No clear effect |
| use tNTK instead of NTK | conditional Monge gap AL · K562-Chemical and BMDM-Genetic | change in Sinkhorn divergence: BMDM 20+10: -130.32 -> -144.90 | Improved |
| use tNTK instead of NTK at small budget | conditional Monge gap AL · K562-Chemical (10+1) | change in Sinkhorn divergence: -9.33 -> -8.65 | No clear effect |
| use tNTK instead of mean-based NTK | affine Gaussian transport AL · Synthetic Gaussian 2D | Wasserstein loss: 1.2x lower than random | Improved |
| use Wasserstein-risk acquisition instead of mean-only acquisition (ablation) | 1D GP active learning · Synthetic 1D conditional density | Wasserstein risk | Improved |
| use NTK on mean AL instead of random | 1D GP active learning · Synthetic 1D conditional density | shape error | No clear effect |
| use NTK from mean-var or quartile networks instead of mean only | NTK baseline · Quantile datasets | change in Wasserstein test risk: Gaussian mixture 5+1: -416.43 vs -367.99 | No clear effect |
| use BALSA-EMD instead of random | active learning baseline · Dew Point (quantile, 5+1) | change in Wasserstein test risk (x1e-3): -79.90 vs -79.51 | No clear effect |
| use tNTK instead of random | conditional Monge gap AL · K562-Chemical (10+1) | change in Sinkhorn divergence: -9.56 -> -8.65 | Hurt |
Experimental setup
Builds on: neural tangent kernel
Closest papers by meaning
- Generative Conditional Distributions by Neural (Entropic) Optimal Transport (ICML 2024 · similarity 0.87)
- A Likelihood Based Approach to Distribution Regression Using Conditional Deep Generative Models (ICML 2025 · similarity 0.86)
- BridgeTwist: Twisting Schrödinger Bridges for Training-Free Conditional Sampling (NeurIPS 2026 · similarity 0.85)
- Active Flow Expansion for Out-of-Distribution Discovery: from Theory to Molecules (NeurIPS 2026 · similarity 0.85)
- Posterior Inference in Latent Space for Scalable Constrained Black-box Optimization (NeurIPS 2026 · similarity 0.85)
- Zero-Shot Conditioning of Score-Based Diffusion Models by Neuro-Symbolic Constraints (AAAI 2025 · similarity 0.84)
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.
