{"paper":{"title":"Post-Training Augmentation Invariance","license":"http://creativecommons.org/licenses/by/4.0/","headline":"Lightweight adapter networks trained with Wasserstein-based losses can add approximate invariance to augmentations in a frozen pretrained network while preserving its original behavior.","cross_cats":["stat.ML"],"primary_cat":"cs.LG","authors_text":"Keenan Eikenberry, Lizuo Liu, Yoonsang Lee","submitted_at":"2025-05-16T21:11:51Z","abstract_excerpt":"This work develops a framework for post-training augmentation invariance, in which our goal is to add invariance properties to a pretrained network without altering its behavior on the original, non-augmented input distribution. We define this notion precisely and additionally introduce augmented encoders, which are probabilistic encoders that formalize augmentation-based encoding processes and that serve as our fundamental object of study. We introduce two losses for augmented encoders, namely, Markov-Wasserstein minimization and Wasserstein correlation maximization, and we demonstrate empiri"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"Both Markov-Wasserstein minimization and Wasserstein correlation maximization can be used to train lightweight one-hidden-layer MLP adapter networks E_theta that, when appended to the latent space of a pretrained network F, lead to approximate post-training augmentation invariance, as evidenced by achieving 94% classification accuracy on arbitrarily rotated STL10 images (vs 71% without adapter) and 86% on noisy images (vs 58%) with F frozen and little corruption to original features.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That the proposed losses can enforce invariance on augmented inputs while keeping the adapter nearly isometric on the non-augmented latent distribution of F, which is presented as an empirical outcome but depends on the probabilistic definition of augmented encoders and the specific training procedure for E_theta.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"Develops post-training augmentation invariance via augmented encoders and two Wasserstein-based losses to train lightweight MLP adapters that boost robustness on DINOv2 features for STL10 without fine-tuning.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"Lightweight adapter networks trained with Wasserstein-based losses can add approximate invariance to augmentations in a frozen pretrained network while preserving its original behavior.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"8540854f017fd2f8261ac1490f7d5a7191cc984fea13b1c538c3dc59359521db"},"source":{"id":"2505.11702","kind":"arxiv","version":3},"verdict":{"id":"dc48260b-2116-4d2c-af7d-e1a28e78ae83","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-22T13:59:08.991178Z","strongest_claim":"Both Markov-Wasserstein minimization and Wasserstein correlation maximization can be used to train lightweight one-hidden-layer MLP adapter networks E_theta that, when appended to the latent space of a pretrained network F, lead to approximate post-training augmentation invariance, as evidenced by achieving 94% classification accuracy on arbitrarily rotated STL10 images (vs 71% without adapter) and 86% on noisy images (vs 58%) with F frozen and little corruption to original features.","one_line_summary":"Develops post-training augmentation invariance via augmented encoders and two Wasserstein-based losses to train lightweight MLP adapters that boost robustness on DINOv2 features for STL10 without fine-tuning.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That the proposed losses can enforce invariance on augmented inputs while keeping the adapter nearly isometric on the non-augmented latent distribution of F, which is presented as an empirical outcome but depends on the probabilistic definition of augmented encoders and the specific training procedure for E_theta.","pith_extraction_headline":"Lightweight adapter networks trained with Wasserstein-based losses can add approximate invariance to augmentations in a frozen pretrained network while preserving its original behavior."},"integrity":{"clean":true,"summary":{"advisory":0,"critical":0,"by_detector":{},"informational":0},"endpoint":"/pith/2505.11702/integrity.json","findings":[],"available":true,"detectors_run":[],"snapshot_sha256":"c28c3603d3b5d939e8dc4c7e95fa8dfce3d595e45f758748cecf8e644a296938"},"references":{"count":0,"sample":[],"resolved_work":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57","internal_anchors":0},"formal_canon":{"evidence_count":2,"snapshot_sha256":"c2f43198f9e381e93cfe5a9aff16e34eb6a2b9cd5e8be031007739857446f244"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"}