acceptodds
Withdrawn from ICLR 2027

Test-time adaptation using Transporting Flow

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Test-time adaptation (TTA) considers a pretrained model on source domain and aims to adapt it to target domain without labels, where the target is subject to distribution-shift. Most existing TTA methods are based on entropy-minimization and feature alignment, where the former are classification-specific. While the latter can be extended to regression, they either naively align the entire feature space or use principal component analysis (PCA) to find dimensions with maximum variance and perform alignment along those dimensions. These maximum variance dimensions however may not be those that are important for domain adaptation, as they represent spurious correlations between domains. A natural question then arises as to how we can identify features with maximal geometric displacements if they exhibit low variance. We use Optimal Transport and propose Transporting Flow (TOW), a framework that actively optimizes for a k-dimensional subspace that incurs maximal displacement cost for feature alignment. TOW computes 2-Wasserstein distance on the pre-learned density of source features and the target activations, and leverages Normalizing Flows to implicitly find an optimal transport map that minimizes the transport cost during adaptation. Our proposed method can be extended to both continuous regression and classification tasks in standard settings with any loss function. We show that TOW outperforms baselines on four real-world datasets.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
test-time adaptation, regression, distribution shift, representation

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_opwwOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a test-time adaptation (TTA) method that aim to learn good features by seeking an important subspace. Unlike existing TTA methods that seeks such subspace with maximum variance, the proposed method seeks more informative subspace by using optimal transport (OT) framework with 2-Wasserstein distance. To avoid computational complexity of discrete OT solver, it uses normalize flows (NFs) for approximating OT map. The proposed method can be applied to both classification and regression tasks. Experiments on classification and regression tasks show the effectiveness of the proposed method compared to existing TTA methods.

Critical strengths
  1. TTA is a practically important research topic, and the proposed approach, applicable to both classification and regression tasks, is useful.
  2. Experiments with two classification and two regression tasks show the effectiveness of the proposed method compared to various existing TTA methods.
Critical weaknesses
  1. There are several unclear and problematic points in the proposed method. First, Eq. (5) contains obvious errors: the matrix UkU_k directly acts on the norm, thus, the objective function becomes a matrix (not a scalar), and the argmin operator is not the objective function. In addition, although the proposed method approximates the OT map with NF, it is unclear whether the important constraint that the OT map should satisfy (i.e., the density-preservation constraint described below Eq. 5) can be met by using NF alone. Furthermore, fϕtowf ^{tow} _{\phi} does not appear to be updated in Algorithm 1. However, it should be updated in Eq. 5. The paper also does not provide a concrete procedure for obtaining zisdez ^{sde} _i from the target sample zitz ^t _i with NF fϕsdef ^{sde} _{\phi}, and zisdez ^{sde} _i does not appear in Algorithm 1. More importantly, although learning the subspace UkU _k is a key point in the proposed framework, how to learn it is not described in Algorithm 1. The only explanation is briefly described in Line 209: "We address this by explicitly optimizing UkU _k on the Stiefel manifold with Riemannian updates." Because of these issues, the optimization procedure is difficult to understand and reproduce, raising concerns about the soundness of the proposed method.
  2. There are concerns about whether experimental settings are fair. As described in Lines 278-284, the learnable parameters of the proposed method are different for each dataset. It also appears that the kk values were selected based on the best test performance for each dataset. It is also unclear why different evaluation metrics are used for Camelyon17 and Birdcalls (i.e., the former uses accuracy, but the latter uses F1 score). I would like the authors to justify the validity of these settings.
  3. Unlike standard TTA methods, the proposed method requires additional training for the source NF, which may limit its practical applicability, as it does not directly apply to off-the-shelf classification and regression models.
Questions for authors
  1. (Clarification requests) Please see the comments in the Critical Weaknesses section.
  2. Although the proposed method requires covariate shift-like assumptions for distribution shift, can it handle other distribution shifts, such as concept shift?
Reviewer_JehtOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper addresses the problem that, in feature-alignment test-time adaptation (TTA), the maximum-variance directions selected by PCA are not necessarily the directions most displaced under domain shift. The contributions consist of three parts. (1) It proposes TOW, an optimal-transport-based framework that actively searches for the k-dimensional feature subspace with the maximal 2-Wasserstein displacement cost and performs alignment only within that subspace. (2) It pre-learns the density of source features with a normalizing flow (SDE), so that adaptation requires neither source data nor distributional assumptions on the features. The subspace is optimized with Riemannian updates on the Stiefel manifold, which avoids unstable gradients through eigendecomposition. (3) The method is evaluated on two regression tasks (LIBS spectra and California Housing) and two classification tasks (Camelyon17 and Birdcalls). Experiments show that TOW outperforms all baselines on the regression tasks, but on Camelyon17 its average accuracy is lower than that of PL and SHOT.

Critical strengths
  1. The motivating question is well posed, since low-variance but highly displaced features are indeed a blind spot of PCA-based methods, and the failure of SSA on LIBS supports this point.
  2. The gains on the regression tasks are substantial and consistent, achieving the best results on nearly all LIBS oxides with smaller standard deviations.
Critical weaknesses
  1. The abstract claims consistent and significant improvements, but on Camelyon17 the average accuracy is lower than that of PL and SHOT, the h4 hospital lags far behind, and the gain on Birdcalls is marginal, so the classification results do not support the claim.
  2. The subspace dimension k requires per-dataset manual search and is highly sensitive, as shown by h2 collapsing from 90.3 to 75.0 when k increases to 90. No label-free criterion for choosing k is provided, and the computational overhead is not reported, which is inconsistent with the scalability claim.
Questions for authors
  1. Does the collapse on h4, together with the cases where AdaBN outperforms TOW in the semi-synthetic experiments, suggest that the maximally displaced subspace may select spurious directions under certain shift types?
  2. In a deployment scenario without target labels, how should k and the learning rates be determined in practice?
Reviewer_DknNOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Transporting Flow (TOW), a test-time adaptation framework based on optimal transport. Instead of aligning the entire feature space or selecting high-variance principal components, TOW searches for a kk-dimensional subspace with large geometric displacement between source and target features. A normalizing flow is used to estimate the source feature density and to learn an invertible map that transports target features toward the source distribution. The method is designed for both classification and continuous regression and is evaluated on LIBS, California Housing, Camelyon17, and Birdcalls. The experiments show consistent improvements over several feature-alignment baselines on the regression tasks, although the advantage is less consistent on Camelyon17.

Critical strengths
  1. The paper addresses an underexplored problem: source-free test-time adaptation for continuous regression. Most common TTA methods rely on entropy minimization and are therefore difficult to apply to scalar-valued regression.

  2. The motivation for using Wasserstein geometry is reasonable. Maximum-variance directions from PCA do not necessarily correspond to the directions with the largest domain shift, whereas the proposed objective explicitly targets geometric displacement between source and target feature distributions.

  3. The framework is evaluated across different modalities and tasks, including spectroscopy regression, tabular regression, pathology image classification, and bird-call classification. The LIBS experiments are particularly relevant because they represent a real scientific application rather than a standard image benchmark.

Critical weaknesses
  1. TOW uses different values of kk across datasets, selects k=77k=77 for Camelyon17 after searching a range of candidates, and varies which feature-extractor parameters are updated. The paper does not fully explain whether these choices are made using only unlabeled target data or whether target-domain performance is consulted. Clarifying this procedure would make it easier to assess compliance with the standard label-free TTA setting.

  2. TOW performs strongly on the LIBS and California Housing regression tasks, but its average accuracy on Camelyon17 (88.46%) is below PL, SHOT, and FR. This does not undermine the main idea, but it suggests that the paper’s conclusions may be more convincing if TOW is presented primarily as a promising approach for measurement-shift regression, rather than as a uniformly superior method across both regression and classification.

  3. TOW requires a source density estimator trained on source-domain feature activations, together with a transporting flow optimized during adaptation. Although raw source samples are not accessed at test time, the paper does not report the parameter count, storage cost, pretraining time, adaptation latency, or peak memory of these components. Providing these measurements would enable a more complete comparison with source-statistics and subspace-based methods.

  4. Previous work has studied feature restoration and subspace alignment for source-free regression adaptation, while SSA explicitly selects a low-dimensional PCA subspace for regression TTA [1–3]. The main contribution of TOW appears to lie in combining a displacement-oriented subspace objective with a normalizing-flow implementation for TTA.  [1] Chen, X. et al. Representation Subspace Distance for Domain Adaptation Regression. Proceedings of the 38th International Conference on Machine Learning (ICML), 2021. [2] Eastwood, C. et al. Source-Free Adaptation to Measurement Shift via Bottom-Up Feature Restoration. arXiv:2107.05446, 2022. [3] Adachi, K. et al. Test-Time Adaptation for Regression by Subspace Alignment. International Conference on Learning Representations (ICLR), 2025.

Questions for authors

Refer to the weakness

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.