acceptodds
Under review as a conference paper at ICLR 2027

Transferable Rankings, Shifted Action Boundaries: The Value of BatchNorm Memory

Abstract

BatchNorm (BN) running statistics retain information from earlier domains, but when should a model carry that state into a new domain? We measure its value through paired carry/reset interventions with frozen weights and identical target batches. Across 12 CIFAR-C receiver checkpoints and all 210 directed corruption transitions per checkpoint, an external effect map correlates with receiver effects (, ), yet its zero-threshold rule loses 1.44 percentage points relative to resetting and is harmful at every checkpoint. On a locked 150/60 edge split, a single threshold learned from the other 11 receivers gains 0.83 points on held receivers and held edges, close to the 0.86 points from a receiver-specific affine fit; the zero threshold loses 1.02 points on the same edges. The transferred threshold also gains 0.93 points over resetting on 120 closed-loop streams with known domain identities and boundaries, without using held-receiver labels. State and layer interventions show that predictive utility is nonlinear even though BN-buffer differences decay exponentially under matched updates. These results identify a shared shift in the useful carry/reset threshold across the tested receivers: much of the source ranking remains useful once the threshold is calibrated.

Primary area:
transfer learning, meta learning, and lifelong learning
Keywords:
test-time adaptation, batch normalization, continual adaptation, distribution shift, memory retention, cross-system transfer

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_6mHdOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
3(+1)
Presentation
3(+1)
Soundness
2
Read review
Paper summary

This paper investigates whether BatchNorm running statistics should be carried or reset across domain shifts. Paired interventions across 12 CIFAR-C receiver checkpoints show that source effect rankings transfer, but their decision boundaries do not. A threshold calibrated on other receivers improves held-edge decisions and closed-loop performance.

Critical strengths
  1. The central finding is insightful. The paper distinguishes ranking transfer from decision-boundary transfer, demonstrating that high correlation does not guarantee beneficial carry/reset decisions.

  2. The empirical controls are carefully designed. Held-receiver and held-edge tests support threshold transfer, while a separate closed-loop evaluation reports positive gains across all 120 streams, which also include profile edges.

Critical weaknesses
  1. The generalizability remains limited. The experiments use frozen convolutional networks, known domain boundaries, and labeled calibration effects from other receivers. Moreover, leave-one-receiver-out testing retains models with the same dataset and architecture, leaving group-level transfer insufficiently tested.

  2. The practical gains are modest. Closed-loop improvement is approximately 0.93 percentage points, and the evaluation includes profile transitions. Comparisons with standard batch-statistics adaptation and different BN momentum settings are needed to clarify the practical value of the proposed decision rule.

Questions for authors
  1. Does the transferred threshold remain effective when an entire dataset–architecture group is held out, rather than a single receiver checkpoint?

  2. How does the threshold policy compare in absolute accuracy with batch-statistics adaptation and alternative BN momentum settings?

Reviewer_QLPNOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates when BatchNorm (BN) running statistics should be carried over or reset during cross-domain adaptation. Through paired interventions with frozen weights and identical target batches, the authors find that effect rankings transfer across models, whereas directly applying a zero-threshold carry/reset rule can be harmful. A shared threshold calibrated on other receivers yields positive gains on held-out receivers, held-out transitions, and closed-loop streams. State interpolation and layer-wise interventions further demonstrate that persistence of BN state does not necessarily translate into sustained predictive benefit.

Critical strengths
  1. The paper clearly distinguishes the transferability of effect rankings from that of decision rules. Figure 1 demonstrates that strong correlations can coexist with negative decision gains, highlighting why correlation alone is insufficient to assess the practical value of BN memory.

  2. The carry and reset branches share identical weights and target batches. Within each experimental system, acquisition, calibration, and evaluation use disjoint underlying image identities. This design helps isolate the effect of incoming BN state while reducing confounding from other training mechanisms and image reuse.

  3. The shared threshold achieves approximately 0.83 percentage points of improvement on held-out receivers and transitions, and 0.93 percentage points in closed-loop streams where decisions affect subsequent states, without using held-receiver labels for calibration. These results provide direct evidence of decision-level utility.

Critical weaknesses
  1. The core experiments focus on a fixed set of CIFAR-C corruptions at a single severity level. Holding out one receiver does not exclude an entire architecture or dataset, and holding out transitions does not test unseen corruption types. Consequently, the results primarily support transfer within the evaluated receiver collection rather than under broader distribution shifts.

  2. Although the method fits only one threshold, each fit uses 150 labeled transition effects from each of 11 other receivers and assumes known domain identities and boundaries. The closed-loop baselines mainly include always carrying, always resetting, and fixed half-interpolation. Comparisons against well-tuned statistics-interpolation or reset strategies under matched information budgets are missing.

  3. The central transfer analysis relies on a single external source effect map, while the source and receiver systems differ simultaneously in preprocessing and training recipes. Although the interpolation analysis reveals nonlinear predictive responses, it does not establish which factors primarily cause the threshold shift or whether the shared threshold depends on this particular source system.

Questions for authors
  1. Could the authors evaluate transfer with an entire architecture, dataset, or corruption family held out, and examine whether the shared threshold remains effective when BN momentum or batch size changes?

  2. How does decision gain change when fewer receivers or transition edges are available for calibration? Could the authors compare against tuned statistics-interpolation or adaptive reset strategies under the same label and domain-information budgets?

  3. Could the authors use multiple external source models and separately control preprocessing and training recipes to identify the main causes of the threshold shift and assess its stability?

Reviewer_Z1rjOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
3
Read review
Paper summary

The paper is trying to answer this question: "does the pattern of benefiting from carrying batch-norm statistics across domain shifts generalize across models and datasets?" In order to do so, the authors consider one source for assessing the carrying capabilities, namely a WRN-40-2 trained on CIFAR-10, against 4 different targets, namely ResNet-20 and WRN-40-2 trained on either CIFAR-10 or CIFAR-100 for 3 seeds. They show that the effects do generally transfer across models, although the source model does consistently overestimate the benefits, with the margin of improvement remaining modest. Always-reset is within 1.5 percentage points of an oracle, and the transferred threshold gains 0.8–0.9 percentage points over it.

The paper is very difficult to read, and it is not related to my field of study. And I admit, I was not able to understand the entirety of the content. Consequently, I reduced my confidence score for the review.

Critical strengths
  • Within its small experimental setting, the paper is rigorous, and the authors scope their claims to the tested source map and receiver family (section 7). I consider the experiments sound. However, my concern is that the scope is narrow (see weaknesses).
Critical weaknesses
  • The paper is almost unreadable. I required some AI agent assistance to understand most of the paper. It uses complicated, jargon-heavy language that overloads certain phrases (eg. calibration) that are common within the ML context. This makes reading it even more difficult. The terms such as receiver, donor, cell, cell–edge, profile/held edges, geometry pool, calibration-role/evaluation-role, action boundary, external source map, and so on are used but never clearly explained. The way the paper presents some of the numerical data, such as in lines 203-212, section 4.2, and lines 305-315 makes the writing confusing, especially paired with the unidentified terminology. It's because the authors list many numbers before clarifying what each represents.
  • The figures and plots are not informative. The legends are often confusing. For example, in Figure 2 R20 and R50 denoting the two models in the legend is bit confusing, and the dataset is not denoted. Furthermore, in Figure 3 the runs are denoted by configurations that are not clarified in the paper.
  • Considering the small margins the paper is dealing with, I am not really convinced about the relevance of the question.
  • The scope of the experiments is narrow. One source model, and four dataset/architecture combinations. Lack of more sophisticated baselines.
Questions for authors

I would like to ask the authors to provide a plain language summary about what the paper's claims are, what each section is trying to do, and a glossary of what each term in the paper means (eg. receiver, donor, cell–edge, calibration-role, action boundary).

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.