Research map / Label-efficient and robust learning / Recognition and long-tail

Show Me What You Don’t Know: Efficient Sampling from Invariant Sets for Model Validation

NeurIPS 2026 · Label-efficient and robust learning › Recognition and long-tail · approach: Data augmentation and generalization bias · 0 citations (OpenAlex)

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 training-free invariance sampling by guiding pretrained diffusion/flow-matching models with a fiber loss via NDTM, plus fiber loss as regularizer for conditional fiber models, to audit feature extractors.

Problem

Analyzing what a feature extractor is invariant to requires sampling from its fibers (inputs mapping to the same features), but existing conditional generative approaches need a new fiber model trained from scratch for each subject model and offer no control over invariance level.

Approach

Use a pretrained unconditional diffusion/flow-matching model as a prior and guide its denoising with the fiber loss on Tweedie's estimate (NDTM), with a modified terminal loss that re-noises the input image to match detail level; also use fiber loss as a regularizer when training conditional fiber models.</af>

Key results

• On ImageNet, NDTM fiber samples for DINOv2 reach fiber loss 273 vs 1225 for nearest neighbors, with TF FID 28.34 vs 26.21 baseline.
• On cue conflict, shape label recovery under ConvNeXt-L rises from 25% on originals to 35% on DINOv2 fiber samples.
• NDTM training-free sampling takes about 1.8 min per sample versus about 14000 min to train a conditional model, breaking even only beyond ~9k samples.

Surprising findings

• Inception fibers yield a PyTorch FID of 3.81 vs TF FID of 33.52, showing FID can be artificially low when generator errors align with the metric network's fibers.
• The Qwen-2B vision encoder places situs inversus images in the same fiber as typical anatomy, suggesting heart laterality is weakly encoded or discarded.

Limitations

• Guidance sampling is slower per sample than conditional model sampling, so it is only cheaper for fewer samples.
• Conditional diffusion/flow models achieve better consistency but are expensive to train with fiber loss.
• Sample refinement in data space leads to pixel-level adversarial-like changes unless done in a latent space.

Experimental setup

Datasets: ImageNet CheXpert MNIST Cue Conflict

Benchmarks: ImageNet CheXpert Cue Conflict Color MNIST

Metrics: fiber loss FID KL-divergence top-5 accuracy AUROC label preference

Baselines: Conditional normalizing flows Free-form flows Free-form injective flows Denoising normalizing flows Multilevel flows Conditional diffusion/flow matching NDTM w/ OOD

Builds on: NDTM conditional normalizing flows free-form flows denoising normalizing flows Tweedie's formula

Closest papers by meaning

• SMI: Statistical Membership Inference for Reliable Unlearned Model Auditing (NeurIPS 2026 · similarity 0.84)
• Diagnosing and Rectifying Fake OOD Invariance: A Restructured Causal Approach (AAAI 2024 · similarity 0.84)
• Is Cross-validation the Gold Standard to Estimate Out-of-sample Model Performance? (NeurIPS 2024 · similarity 0.83)
• From Bricks to Bridges: Product of Invariances to Enhance Latent Space Communication (ICLR 2024 · similarity 0.83)
• Learning Conditional Invariances through Non-Commutativity (ICLR 2024 · similarity 0.83)
• Leveraging Data Symmetries to Select an Optimal Subset of Training Data under Label Noise (NeurIPS 2026 · similarity 0.83)