acceptodds
Under review as a conference paper at ICLR 2027

BIDETA: Brain-Inspired Data-Efficient Tactile Adaptation for Unseen Sensors

Abstract

Vision-based tactile sensors provide high-resolution contact information for robotic perception and contact-rich manipulation, advancing embodied intelligence through more reliable physical interaction. However, device-specific sensing mechanisms cause tactile foundation models to degrade on unfamiliar hardware. Existing cross-sensor methods often require calibration data, paired observations, or iterative training. To address this problem, we propose Brain-Inspired Data-Efficient Tactile Adaptation (BIDETA), a gradient-free framework that uses a frozen tactile encoder and a few labeled target contacts to jointly predict labels for an unlabeled query batch. Inspired by the brain’s rapid sensory adaptation, BIDETA combines rapid support memory, support-conditioned spectral graphs, and reliability-gated recurrence to preserve pretrained representations, repair sensor-dependent feature neighborhoods, and integrate reliable cross-query evidence. Experiments on SITR, TacVerse Shape, and TacQuad show that BIDETA substantially improves adaptation to unknown sensors: with only 10% labeled target data on SITR, it raises mean Sparsh accuracy from 6.86% for the frozen source classifier to 87.09%, exceeding the strongest implemented prior comparison by 47.22 percentage points, and these gains generalize across datasets, pretrained backbones, and tactile tasks. In the SITR timing benchmark with TVL, BIDETA also achieves approximately 20× faster target-sensor adaptation than the best baseline. BIDETA thus offers a gradient-free, data-efficient route to deploying tactile models on new hardware.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Tactile Adaptation, Data-Efficient, Brain-Inspired Learning

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_1iEnOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents BIDETA, a framework for adapting frozen tactile encoders to unfamiliar sensors using a small labeled support set and an unlabeled query batch. The method combines statistical classifiers fitted to two encoder readouts, a graph adjusted using variation in the support examples, and recurrent prediction updates controlled by uncertainty. Using TVL and Sparsh, the authors then evaluate classification and identity ranking tasks across three datasets.

Critical strengths
  1. Being able to adapt trained encoders to new tactile hardware with limited labeled data is a useful practical problem. The combination of support memory, spectral adjustment of feature geometry, and graph propagation is justified well; keeping the encoder frozen also makes the approach attractive when retraining is costly.

  2. The evaluation across SITR, TacVerse Shape, and TacQuad, with two different pretrained encoders, is a strong point. The classification gains are substantial across most sensors and label settings. I also appreciate the identity ranking results, which provide evidence beyond classification.

  3. On SITR with TVL at 10% support, adding the graph and recurrence to the memory improves accuracy from 76.07% to 85.71%. This result really helps understanding their combined contribution.

  4. The appendix is very detailed and informative as it includes accuracy and macro F1 by sensor, results with 2, 5, and 10 examples per class, confusion matrices, both ranking metrics, and runtime measurements. Together with the parameter studies and results over three seeds, this makes the experimental evidence much easier to assess.

Critical weaknesses
  1. The ablations explain the gains within BIDETA, but the external comparisons use different feature readouts (I see that Table 6 gives BIDETA two readouts while each baseline uses one). I would like to understand how these choices were made and whether giving the baselines access to the same features would change the comparison.

  2. I am less convinced by the benefit of the full reliability gate than by spectral normalization: in Table 4, using entropy or disagreement alone gives nearly the same results as the full gate. A comparison with a fixed recurrence weight would help show how much is gained by adjusting the weight for each query.

  3. The gains are less consistent with very few support examples. For example, BIDETA with Sparsh falls slightly below the strongest comparison on TacTip and ViTac at 1% support. The paper does suggest that the three examples per class may be unrepresentative, but I’m not sure whether the main difficulty is already present in the support memory or is amplified by graph propagation.

Questions for authors
  1. How did you choose the feature readout for each baseline? Would the strongest comparisons improve if they could use both readouts?

  2. How does BIDETA behave with uneven class counts or a smaller query batch? Does matching the query class proportions to the support set still help in those cases?

  3. Since the simpler gates are already very close to the full model, would a fixed recurrence weight work just as well?

  4. For the small drops on TacTip and ViTac with 1% support, can you tell whether the graph step helps or hurts the memory predictions?

3 comments

We thank the reviewer for recognizing the practical value of our framework and the breadth of our experimental evaluation. We address the comments and questions in order below.

Weakness

1. The baselines use different feature readouts.

Please see our response to Question 1.

2. The full reliability gate improves prediction.

We believe that the similar results in Table 4 do not negate the contribution of the full gate. Single-signal variants still adjust propagation weights for individual queries. These ablations therefore primarily compare reliability signals and do not establish that a fixed weight can replace query-specific gating.

To test this directly, we fixed the LDA predictions, spectrally corrected graph, and class-prior normalization on SITR with Sparsh and 10% labeled data. Assigning all queries the original gate’s mean weight yields 86.12 ± 0.69% accuracy. Shuffling only the correspondence between weights and queries yields 86.02 ± 0.67%. Full BIDETA achieves 87.09 ± 0.61%, improving accuracy by 0.97 and 1.07 percentage points, respectively. These results show that assigning weights to individual queries improves prediction at the same mean propagation strength, supporting the contribution of the full gate.

3. Causes of difficulty with very few support examples

On these two sensors, the difficulty is already present at the support memory stage, while graph propagation remains beneficial on average across three random seeds.

To distinguish the effects of support memory and graph propagation, we additionally analyzed predictions on TacVerse Shape with Sparsh and 1% labeled data. Using identical support sets, query sets, and class-prior normalization, the mean accuracies without graph propagation are 38.44% and 58.52% on TacTip and ViTac, respectively. Adding graph propagation increases them to 39.00% and 58.74%, yielding gains of 0.56 and 0.22 percentage points. These results show that graph propagation still improves memory predictions on average with only three support examples per class.

Question

1. Does the comparison hold when baselines use the same dual readouts?

Some baselines improve with dual readouts, but BIDETA retains its advantage. The original experiments used predefined readout configurations in our baseline implementations. Methods relying on the source classifier used the final readout to match its input, while SimpleShot and LaplacianShot used the earlier readout to construct class prototypes and query neighborhoods.

To directly examine the effect of readout selection, we conducted additional comparisons using identical frozen dual readouts, support samples, and query splits. We provided the concatenated readouts to the baselines and refitted the corresponding source classifiers for the three adapters while retaining their architectures and training objectives. LaplacianShot was retuned using internal validation on the Mini_3 training set and achieved 39.73% on SITR with Sparsh and 10% labeled data, close to its original 39.87%.

The dual-readout adapter results below report accuracy in percentages as mean ± standard deviation over three random seeds with 10% labeled data.

MethodSITR TVLSITR SparshTacVerse TVLTacVerse Sparsh
SITR-Calib47.50 ± 0.6332.91 ± 1.38N/AN/A
AnyTouch Match45.91 ± 0.5234.86 ± 0.3256.84 ± 1.0749.44 ± 1.25
CTSRL CSM44.91 ± 0.7836.92 ± 0.5459.41 ± 2.2848.98 ± 1.62
BIDETA85.71 ± 0.1287.09 ± 0.6179.13 ± 1.2083.58 ± 1.05

SITR and TacVerse results average six and five target sensors, respectively. These results show that differences in readout count are insufficient to explain BIDETA’s performance advantage. We will supplement Table 6 and Appendix A.4 with the matched-readout results and implementation details.

2. Does prior matching still help with class imbalance and smaller query batches?

In our additional experiments with imbalanced support class counts and smaller query batches, prior matching still improves mean accuracy. We evaluated SITR with Sparsh and 10% labeled data. The following accuracies are averaged over six target sensors and three random seeds.

SettingDual-readout LDA (%)Without prior normalization (%)Full BIDETA (%)
Support class counts with a maximum-to-minimum ratio of approximately 3 to 1, balanced queries77.2681.6781.78
Query batch size 6477.9276.2378.41
Query batch size 25677.9277.3680.02

The imbalanced-support experiment keeps the total labeling budget unchanged, and full BIDETA outperforms dual-readout LDA by 4.52 percentage points. With query batch sizes of 64 and 256, prior matching improves accuracy by 2.18 and 2.66 percentage points, respectively, showing that its benefits persist with smaller query batches.

3. Can a fixed recurrence weight achieve the same performance?

As discussed in our response to Weakness 2, a fixed weight does not fully retain the gains from query-specific gating at the same mean propagation strength. On SITR with Sparsh and 10% labeled data, keeping all other configurations identical and assigning every query the original gate’s mean weight yields 86.12 ± 0.69% accuracy, compared with 87.09 ± 0.61% for full BIDETA. Query-specific gating improves accuracy by 0.97 percentage points, supporting the benefit of adjusting propagation strength for individual queries.

4. Does graph propagation improve memory predictions?

Graph propagation remains beneficial on average on both sensors. As discussed in our response to Weakness 3, on TacVerse Shape with Sparsh and 1% labeled data, keeping the support sets, query sets, and class-prior normalization identical, adding graph propagation increases mean accuracy from 38.44% to 39.00% on TacTip and from 58.52% to 58.74% on ViTac, yielding gains of 0.56 and 0.22 percentage points, respectively. These results are averaged over three random seeds and support the benefit of graph propagation for memory predictions.

Reviewer_1iEnOpenReview ↗

Post Rebuttal Comment

Thanks for the extra experiments and work. The comparisons using the same feature readouts addres my concerns about the baselines. While the gain over the fixed mean weight is still pretty modest, it provides some evidnce for query-specific gating. Please include these results in the final version and make clear how the baselines were adapted and the fixed weight was chosen. I still think this approach can be of interest for the community, so I’ll keep my rating unchanged.

Reviewer_7dC8OpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper proposes BIDETA, a training-free transductive adaptation pipeline for frozen tactile encoders (TVL, Sparsh) on unseen tactile sensors. For each image, it extracts two feature vectors, fits Shrinkage LDA classifiers on the support set, whitens features using Ledoit-Wolf within-class support covariance to build a mutual kk-NN query graph, and runs label propagation gated by entropy and cross-readout Jensen-Shannon divergence (RR), applying Sinkhorn-Knopp column/row scaling at every iteration.

Critical strengths

None

Critical weaknesses
  1. The paper confounds the core of it's methods with neurological terminologies that adds little to the clarity of the method, and instead makes it very difficult to understand the core of the contribution. At the current state the paper's presentation is far from acceptable.
  2. Many terminology and symbols are not well-defined in the paper, such as the two readouts h(1)(x)h^{(1)}(x) and h(2)(x)h^{(2)}(x). Experiment designs are questionable, such as the SITR experiments setting λ=1.0\lambda = 1.0 which makes the entropy term ablation obsolete.
Questions for authors

Are you able to re-explain your method in a clear and precise term without the involvement of unnecessary terms and emphasize the actual important details?

4 comments

Thank you for reviewing our work. We believe some aspects of our method may have been misunderstood, and we hope our responses below clarify these points and address your concerns.

Weaknesses

1. Content conflation

We respect your opinion but disagree that the paper conflates brain-inspired motivation with the core method.

The brain-inspired motivation explains the origins of our algorithmic design, while the algorithm itself is presented through self-contained mathematical definitions and computational steps. The introduction discusses the research significance, related work, problems to be addressed, and connections between brain mechanisms and algorithm design. The method section presents the motivation in separate paragraphs and explicitly defines how support samples produce class probabilities, how query geometry is corrected, and how prediction reliability controls recurrent updates. Readers can understand the complete algorithmic pipeline from these definitions even if they skip the neuroscience background.

This organization is intended to accommodate readers from different fields. Readers interested in neuroscience can explore the design motivation, while those interested in computational methods can directly follow the algorithmic definitions and derivations without prior neuroscience knowledge.

2. Terminology, notation, and gate configurations

We respect your opinion but disagree with the assessment that many terms and symbols are not clearly defined in the paper.

Section 3.1 defines h(v)(x)∈Rdvh^{(v)}(x)\in\mathbb{R}^{d_v} as the vv-th feature readout of the frozen encoder, where v∈{1,2}v\in\lbrace 1,2\rbrace and dvd_v denotes its dimensionality. Appendix A.3 and the footnote of Table 4 further specify that TVL uses global average-pooled and projected features, while Sparsh uses mean-pooled normalized patch tokens from the penultimate and final blocks. Table 6 lists whether each method uses one or both readouts. Reviewers ozxj and 1iEn also explicitly discuss the distinction between single-readout baselines and dual-readout BIDETA.

The disagreement weight is selected through internal validation on the Mini_3 training set of SITR, and TacVerse directly reuses the corresponding configurations. We additionally evaluated the following supplementary configurations.

Label budgetTVL disagreement weightSparsh disagreement weight
1%1.01.0
5%0.500.50
10%1.00.75

When λ=1.0\lambda=1.0, the entropy term is indeed inactive, and the full gate and disagreement-only gate produce identical predictions, consistent with Equation (9). The entropy-only variant uses λ=0\lambda=0 to compare different reliability signals, so this comparison remains meaningful.

For the six classification settings that originally used λ=1.0\lambda=1.0, we additionally conducted a test-set grid sensitivity analysis. Selecting the best non-1.0 candidate for each setting yields a maximum improvement of 0.1333 percentage points and a mean improvement of 0.0291 percentage points over 1.0. This is consistent with our intended protocol of selecting suitable λ\lambda values through validation rather than directly choosing the best test-set values. λ=1.0\lambda=1.0 is one of the values selected through validation for particular settings. This supplementary analysis does not participate in main-experiment parameter selection.

Table 26 further reports internal-validation results for different disagreement weights at the 5% label budget. We will revise Table 7 to distinguish the actual main-experiment parameters from supplementary configurations and specify the parameters for each backbone and label budget.

Question

1. How does BIDETA work, and what is its core contribution?

The abstract and introduction clearly present the research problem, core design, and main experimental results. Section 3 and the method diagram explain the computations and connections between components, while the appendix pseudocode and Section A.3 provide the computational steps and specific feature-readout definitions. These descriptions progress from the overall idea to mathematical definitions and implementation details, helping readers with different backgrounds and levels of familiarity understand the method.

Specifically, BIDETA uses a small labeled support set from the target sensor to calibrate both class predictions and query neighborhoods while retaining the class evidence provided by the support samples. Its inputs are a frozen encoder, a labeled support set, and an unlabeled query batch. Adaptation neither updates the encoder nor uses query labels.

  1. Construct initial class predictions. Extract two feature readouts from the same encoder, standardize them using support-set statistics, fit a shrinkage LDA classifier for each readout, and fuse their class probabilities into the fixed initial prediction matrix P0P_0.

  2. Adjust query neighborhoods. Estimate within-class covariance from the residuals of support features around their respective class means, and downweight directions with large within-class variation. Construct nearest-neighbor graphs using the transformed features and fuse neighborhood connections supported by both readouts to obtain the graph matrix GG. This transformation affects only graph construction and leaves the initial class probabilities P0P_0 unchanged.

  3. Control propagation according to query reliability. Compute a propagation coefficient for each query from a weighted combination of initial prediction entropy and disagreement between the two readout predictions, forming a diagonal matrix RR. Starting from P(0)=P0P^{(0)}=P_0, update predictions as follows.

P(t+1)=Nπ[(I−R)P0+RGP(t)].P^{(t+1)} = \mathcal{N}_{\pi}\left[(I-R)P_0 + RGP^{(t)}\right].

Here, II is the identity matrix. The first term retains the initial class probabilities, while the second aggregates neighborhood predictions. The operator Nπ\mathcal{N}_{\pi} performs alternating column and row normalization using the support-set class frequencies π\pi. After the iterations, the final class probabilities determine classification predictions or identity rankings.

The core contribution is that the target support set provides evidence for class discrimination and corrects query geometry, while propagation consistently retains the initial class evidence.

As reported in the abstract and introduction, with only 10% labeled target data and the encoder frozen, BIDETA increases TVL accuracy on SITR from 7.58% to 85.71%, outperforming the strongest external comparison by 38.24 percentage points. On TacVerse Shape, Sparsh accuracy increases from 13.67% to 83.58%, outperforming the strongest external comparison by 27.60 percentage points. These results demonstrate the practical value of adapting to unseen tactile sensors with limited labeled data across different pretrained encoders and datasets.

After examining your plain-language summary alongside the additional rebuttal controls, the underlying pipeline is much easier to follow, and I recognize that meaningful effort went into the empirical design and cross-sensor evaluations. Simply put, had the concise 1–2–3 statistical formulation from your rebuttal been the core presentation in Section 3, the manuscript would be substantially stronger and far easier to evaluate than the submitted version. Given the additional explanation, I will also consider my confusion on the readout definition resolved.

That being said, the connection to neurological mechanisms operates only at a loose, metaphorical level and actively confounds the manuscript rather than clarifying it. Throughout the paper, standard statistical operations are renamed without adding technical meaning—for instance, "rapid support memory" is simply target-feature z-score standardization followed by Shrinkage LDA which is a common techinique in the adaptation community. Coupled with omitted per-budget hyperparameters in the main text (e.g., Table 7 vs. the actual λ\lambda, ρ\rho, τm\tau_m, and α\alpha settings), I stand by my assessment that the current writing and framing do not meet the bar for acceptance.

Beyond terminology, the manuscript still lacks a scientific narrative on three fronts:

Motivation: Out-of-distribution generalization and few-shot transfer across sensors are well-studied problems. The paper does not clearly articulate what specific, unaddressed technical bottleneck in cross-sensor tactile transfer requires a new framework beyond standard low-shot domain adaptation.

Methodological Novelty: Because each mathematical building block is well established, it remains unclear which specific part constitutes a meaningful algorithmic advance over existing tactile adaptation or transductive few-shot methods, especially when components like the dual-signal reliability gate yield negligible gains over simpler constants.

Experimentation: As also highlighted by Reviewer ozxj, the main experimental tables in the paper used an unbalanced setup (giving only BIDETA dual readouts, target covariance scaling, and transductive batch priors while comparing against single-readout or inductive baselines). Once standard matched-readout baselines like Concatenated Shrinkage LDA (78.70%) or Linear Probing (73.06%) are included, the margin shrinks significantly, showing that most of the reported jump comes from basic linear/LDA calibration on standardized target features rather than the proposed brain-inspired architecture. Comparison with proper adaptation baselines should be a required and emphasized component in the main paper, and it confounds me why this is only disclosed in the discussion period, which is supposed to be used mostly for clarification rather than building foundational results.

Thank you for your further response. We address your points below. 

We disagree with your negative assessments of the paper’s presentation, motivation, novelty, and experimental evidence. These points are supported by evidence in the paper and our responses, which is inconsistent with your current characterization. 

1. Paper presentation 

As explained in our previous response, the abstract, introduction, and experimental sections present the research problem, methodology, and experimental results. The method section separates the brain-inspired motivation from the algorithmic definitions, allowing readers from different disciplines to follow the aspects relevant to their interests. In your latest response, you also acknowledge that the computational pipeline is easier to understand and consider the readout-definition issue resolved. If you still find these independently specified computational steps difficult to understand, we deeply regret this. 

The hyperparameters are selected through internal validation on the SITR Mini_3 training set, and TacVerse reuses the corresponding configurations. Table 7 summarizes the hyperparameters, while Tables 23–29 and 30–34 report parameter sensitivity for classification and ranking, respectively. Regarding the incomplete per-budget reporting of the LDA shrinkage, memory temperature, and readout-fusion weight in Table 7, we have provided supplementary configurations in our previous responses and committed to aligning the revised table with the actual configurations used. We believe that the existing sensitivity experiments, supplementary configurations, and anonymously released code jointly provide concrete support for understanding parameter effects and independently reproducing the method. 

2. Motivation and novelty 

The introduction already discusses our research question and its importance, namely how a tactile foundation model can use a few labeled target contacts to transfer its perceptual capability rapidly and reliably across unseen sensors. 

The specific technical bottleneck is that sensor changes simultaneously misalign class decision boundaries and distort query neighborhoods. Adjusting a classifier alone does not directly repair relationships between queries, while propagation over distorted neighborhoods may weaken the class evidence provided by support samples. We therefore use the same target support set to establish class evidence and correct query geometry, while anchored propagation continuously preserves the initial class evidence. 

As explained in our response to Reviewer ozxj’s Weakness 2, our contribution lies in unifying established components for label-efficient adaptation to unseen tactile sensors. 

3. Experimental results 

If you refer to Reviewer ozxj’s concerns about readout conditions, the matched experiments we have already provided should also be considered. On SITR with Sparsh and 10% labeled data, using identical frozen dual readouts, support samples, and query splits, with hyperparameters retuned through internal validation, LaplacianShot’s accuracy decreases from 39.87% to 39.73%, remaining close to the original result. 

For the matched classifiers specifically cited in your latest response, concatenated shrinkage LDA achieves 78.70%, the linear classifier achieves 73.06%, and full BIDETA achieves 87.09%, yielding improvements of 8.39 and 14.03 percentage points, respectively. Adaptation through target-class statistics is itself part of our framework. These results further show that the complete method retains additional gains over simple classifiers after matching feature access. We do not know whether omitting these remaining gains when citing the matched comparisons was inadvertent or intended to mislead. We are willing to assume the former. 

Likewise, we disagree with the characterization that evidence of component contributions was disclosed only during the discussion period. Table 4 in the original paper already reports accuracy for different component configurations, and Table 22 provides the corresponding Macro-F1 results. Under the same SITR, Sparsh, and 10% labeling setting, dual memory without a query graph achieves 77.92%, the variant without spectral correction achieves 82.29%, and full BIDETA achieves 87.09%. These results, already present in the paper, support the performance advantage of the complete brain-inspired architecture over the tested simplified configurations. We do not know whether you have forgotten these previously reported results or have not read the relevant tables. We are willing to assume the former. 

We have provided comparative experiments, component ablations, sensitivity analyses, and anonymous code to support our claims. We trust that the AC and readers will make independent, fair judgments based on this evidence.

Reviewer_ozxjOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

BIDETA studies adaptation of frozen tactile encoders to target sensors using limited labeled support data. It combines two shrinkage-LDA classifiers, a support-conditioned spectral transformation for graph construction, and gated propagation over an unlabeled query batch with class-prior normalization. Experiments use TVL and Sparsh on SITR and TacVerse Shape for classification, and TacQuad for closed-set trial-identity ranking. The paper reports improvements across support budgets, with additional few-shot evaluations, component ablations, and timing measurements.

Critical strengths
  1. The reported improvements extend across two encoders and multiple target sensors. Appendix B.3 includes 2/5/10-shot evaluations, and the SITR results in Tables 16–17 exclude the sensor used for hyperparameter development. These experiments support the data-efficiency claim beyond percentage-based label budgets.

  2. At the 10% SITR budget, the full method improves over the dual-memory-only variant by 9.64 and 9.17 percentage points with TVL and Sparsh, respectively. Removing spectral normalization also reduces accuracy in Table 4. These results support the usefulness of the adaptation pipeline beyond fitting an LDA classifier alone, while keeping the encoder frozen.

Critical weaknesses
  1. On SITR with Sparsh and 10% labeled data, the strongest external baseline achieves 39.87%, compared with 77.92% for dual-readout LDA and 87.09% for full BIDETA. This gap warrants checking whether feature selection, classifier configuration, or optimization limits the baselines. Table 6 gives only BIDETA both readouts, and Appendix A.4 provides limited implementation details. These differences do not establish an unfair comparison, but stronger comparisons under matched conditions are needed to substantiate the reported advantage.

  2. BIDETA combines established statistical classification and graph-inference components. Such an integration can be useful, but the brain-inspired motivation does not by itself establish novelty. The ablations support spectral correction more clearly than the particular fusion and gating design: arithmetic fusion and single-signal gates perform similarly to the full method, while the sequential ablations are not fully specified. The added value over simpler combinations of the same components therefore remains unclear.

  3. The class-prior normalization in Equation (11) aligns aggregate query probability mass with the support prior, while all evaluations use balanced support and query sets. Without isolating this operation or testing mismatched class proportions and smaller batches, its contribution and deployment implications remain uncertain. Task coverage is also limited: TacQuad ranks enrolled trial identities, and support and query frames can share a physical trial. This supports closed-set ranking, but provides limited evidence for the broader task-generalization claim in Section 4.3.

Questions for authors
  1. Using identical feature readouts, preprocessing, support sets, and validation procedures, how do linear classifiers, LDA, and an MLP compare with BIDETA? Last-block fine-tuning or LoRA under the same label budget, with adaptation costs reported, would provide a practical reference. A representative frozen DINOv2 or DINOv3 with a simple classifier would additionally help assess the role of the pretrained representation.

  2. Keeping the LDA predictions and class-prior normalization fixed, how does BIDETA compare with conventional label propagation and a validation-selected constant propagation coefficient? Please also specify which operations change and which remain fixed across the sequential variants in Table 4.

  3. How much of the improvement remains after removing the class-prior normalization step? With the support set fixed, do the conclusions hold when query-class frequencies are unequal, query batches are smaller, or some supported classes are absent from a query batch?

  4. Which disagreement weight produced the main classification results: 1.0 in Table 7 or the selected value of 0.5 in Table 26? If the weight is 1.0, the entropy term is inactive. Please explain any differences between the full-gate and disagreement-only results and whether their remaining configurations are identical. Please also specify the values or selection rules for the LDA shrinkage parameter, memory temperature, and readout-fusion weight.

  5. For SITR-Calib, AnyTouch Match, and CTSRL CSM, please specify the architectures, objectives, loss weights, learning rates, and tuning data, and explain which original mechanisms were retained or modified. Do the main comparisons remain valid when these methods receive the same two readouts as BIDETA?

  6. Minor Comments: a) In Figure 1, standardize “Tactip” and “ViTactip” to “TacTip” and “ViTacTip,” consistent with the main text. b) Table 29 reports 92.38 ± 2.08 for TVL at the selected iteration count of 10, whereas selected configurations in Tables 23–28 repeatedly report 92.38 ± 1.91. Please check whether this reflects different runs or configurations, or a reporting error.

5 comments

We thank the reviewer for recognizing BIDETA’s data efficiency across encoders and sensors and the benefits of its adaptation pipeline beyond LDA alone. We address each concern below.

Weaknesses

1.Baseline verification and reproducibility

Our dual-readout LDA builds statistical memory from the target support set’s class statistics to adapt to new sensors. Its advantage does not come solely from the model’s two feature readouts. To check whether feature selection or configuration limited baseline performance, we gave LaplacianShot the same two frozen Sparsh readouts, support samples, and query split, and retuned its hyperparameters using validation within the training set. It achieved 39.73% accuracy, close to the original 39.87%.

To address your concerns about implementation details, we provide an anonymous code repository at https://anonymous.4open.science/r/BIDETA-7483/ and subsequently release all code and model parameters publicly to support independent reproduction.

2.Novelty and component contributions

We do not claim that LDA, spectral transforms, or graph propagation are individually new, or that brain inspiration alone establishes novelty. Our novelty lies in how brain-inspired principles unify these components for label-efficient adaptation to unseen tactile sensors, jointly addressing misaligned decision boundaries and distorted query neighborhoods. We use a few labeled target samples to establish a class-probability anchor, correct the query graph using within-class support variation, and preserve the initial class evidence through anchored propagation.

We added controlled experiments to assess these components directly. On SITR with Sparsh and 10% labeled data, keeping the LDA predictions and class prior fixed, standard graph propagation achieves 81.85%, while BIDETA achieves 87.09%, supporting the value of support-conditioned geometry within the adaptation pipeline. For gating, shuffling only the assignment of weights to queries reduces accuracy to 86.02%, supporting query-specific allocation of propagation strength. These results support the contributions of geometry correction and query-dependent gating, without requiring a particular fusion rule or dual-signal combination to be optimal in every setting.

3.Prior normalization and task scope

Our prior normalization does not use query labels at all, relying only on support-set class statistics. On SITR with Sparsh and 10% labeled data, adding the same normalization to dual-readout LDA increases accuracy from 77.92% to 79.39%, still below BIDETA’s 87.09%. Thus, this operation alone does not explain the overall advantage. To further address your concern, we fixed the support set and evaluated BIDETA under a 10-to-1 long-tailed query distribution, achieving 82.62% accuracy compared with 79.57% for dual-readout LDA on the same queries. This shows that performance can be limited under extreme class-proportion mismatch. We will include this limitation in the conclusion. TacQuad evaluates closed-set ranking of enrolled identities on unseen sensors, not generalization to new trials or identities. For each identity, we sample support frames from the first 12 frames, leave a four-frame gap, and use the final four frames as queries. Support and query frames do not overlap, although they may come from the same trial. This experiment assesses the adaptation method’s effectiveness for identity ranking. We will clarify the scope of the conclusions in Section 4.3 accordingly.

Key Questions

1.How does BIDETA compare with other methods?

To address your question, we conducted the additional experiments you requested. On SITR, BIDETA outperforms linear classifiers, LDA, and an MLP under identical feature conditions, as well as last-block fine-tuning under the same label budget. Classifier comparisons use the same frozen dual readouts, preprocessing, support samples, and validation procedure within the training set. The table reports accuracy averaged across six target sensors with 10% labeled data, as the mean and standard deviation over three random seeds.

MethodTVL accuracy (%)Sparsh accuracy (%)
Linear classifier (L-BFGS)71.71 ± 0.3073.06 ± 0.18
Concatenated shrinkage LDA76.56 ± 0.2478.70 ± 0.66
MLP73.77 ± 0.5575.83 ± 0.17
Last-block fine-tuning77.47 ± 1.0476.15 ± 0.99
BIDETA85.71 ± 0.1287.09 ± 0.61

The epoch search for last-block fine-tuning extended to 100 epochs, with selection based on validation within the training set. The median combined times for adaptation and full-batch query inference from cached representations were 30.48s and 0.31s for Sparsh last-block fine-tuning and BIDETA, respectively.

Using the same DINOv2 dual readouts, support set, and query split, LDA and BIDETA achieved 82.83% and 91.32%, respectively, on TacVerse Shape. This shows that BIDETA can further improve predictions with a stronger frozen representation.

2.How does BIDETA compare with simpler propagation?

With the LDA predictions and class-prior normalization fixed, BIDETA achieves 87.09% on SITR with Sparsh and 10% labeled data, compared with 81.85% for conventional label propagation. The support-conditioned spectral graph also achieves performance close to the full model with a validation-selected constant coefficient, showing that geometry correction remains effective under simplified propagation.

To further examine query-wise gating, we added a constant-coefficient control matched to the full gate’s mean weight. It achieves 86.12%, compared with 87.09% for the full gate. Keeping the weight distribution unchanged and shuffling only the assignment of weights to queries reduces accuracy to 86.02%, supporting query-specific allocation of propagation strength.

In Table 4, the dual-memory variant outputs only the fused LDA probabilities. The next three graph-inference variants keep the support samples, LDA probabilities, neighborhood size, graph temperature, iteration count, and class-prior normalization fixed. They construct graphs using raw features, features standardized using the support set, and spectrally corrected features, respectively, all with a constant propagation coefficient of 0.7. Full BIDETA retains the spectral graph and replaces the constant coefficient with query-wise gating ranging from 0.7 to 0.9. We will explicitly document these settings in Table 4 and the appendix.

3.Does the advantage remain without prior normalization or under different query conditions?

The advantage remains after removing class-prior normalization. In additional experiments on SITR with Sparsh and 10% labeled data, BIDETA without this operation achieves 82.13%, compared with 77.92% for dual-readout LDA, retaining an improvement of 4.21 percentage points.

With the support set fixed, the advantage also holds under a 10-to-1 class imbalance. Please see our response to Weakness 3 for the results and their limitations. The average advantage also remains under the tested smaller query batches. With independent batches of 64 and 1024 queries, full BIDETA achieves 78.41% and 84.51%, respectively, both above 77.92% for dual-readout LDA. We construct the graph and run inference independently for each batch without retuning hyperparameters.

In additional experiments where the query set contains only half of the supported classes, BIDETA without class-prior normalization achieves 83.99%, compared with 78.17% for dual-readout LDA, showing that this variant retains adaptation gains in this missing-class setting.

4. How is the disagreement weight set?

We select hyperparameters using internal validation on the SITR Mini_3 training set and directly transfer the SITR configurations to TacVerse. We further evaluate the following disagreement-weight configurations.

Label budgetTVL disagreement weightSparsh disagreement weightLDA shrinkageMemory temperatureReadout fusion weights
1%1.01.00.60 / 0.051.00.60 / 0.40
5%0.500.500.01 / 0.104.00.50 / 0.50
10%1.00.750.01 / 0.011.00.50 / 0.50

We will revise Table 7 to specify the configuration corresponding to each set of results.

When the weight is 1.0, the entropy term is indeed inactive, and the full-gate and disagreement-only variants produce identical per-run predictions, consistent with the formulation. To examine the effect of the disagreement weight, we additionally conduct a grid search on the test sets. Selecting the best non-1.0 weight for each setting yields a maximum improvement of 0.1333 percentage points and an average improvement of 0.0291 percentage points. These results show that 1.0 is not always optimal. It is one of the weights selected through internal validation for particular settings, consistent with our intended design.

5.Does BIDETA retain its advantage with the same two readouts?

Yes. We conducted additional comparisons using identical frozen dual readouts, support samples, and query splits. The two readouts are individually normalized and concatenated with equal weighting, and the corresponding source classifier is refitted. The adapter implementations are specified below.

MethodArchitecture and adaptation of the original mechanismLearning rate
SITR-CalibFollowing the calibration-conditioning idea of SITR, we use a 48-dimensional residual bottleneck, a two-layer 24-dimensional condition encoder, and FiLM modulation. A calibration-difference vector replaces the original joint image and calibration-token encoding.0.002
AnyTouch MatchA 64-dimensional GELU residual bottleneck replaces AnyTouch’s cross-sensor paired matching with cosine alignment between class prototypes.0.002
CTSRL CSMFollowing CTSRL’s sensor-embedding and channel-modulation mechanisms, we use a 32-dimensional compression layer, a 16-dimensional sensor embedding, and a 128-dimensional projector, with correlation regularization applied to class prototypes.0.0015

SITR-Calib constructs its calibration condition from background images and 18 standardized contacts. All three implementations freeze the encoder and perform feature adaptation in place of the original full encoder-pretraining pipelines.

Their objectives share target classification cross-entropy with weight 1.0, source-prototype cross-entropy with weight 0.30, and within-class compactness regularization with weight 0.10. SITR-Calib and AnyTouch Match additionally use prototype alignment with weight 0.35. CTSRL CSM instead adds a Barlow-style correlation loss with weight 0.10 and an off-diagonal coefficient of 0.005.

Learning rates and loss weights remain fixed to those in our original experiments and are not retuned for the dual-readout comparisons. The source classifier uses a learning rate of 0.003, with checkpoints selected on a 10% holdout of the source training set. Target adaptation uses only the support set, AdamW with weight decay 0.0005, at most 300 epochs, and early stopping based on training loss. Query test labels are not used for hyperparameter selection.

Across both backbones and all three label budgets, BIDETA outperforms these dual-readout adapters. Accuracy at 10% labeled data is reported below as mean ± standard deviation over three random seeds.

MethodSITR TVLSITR SparshTacVerse TVLTacVerse Sparsh
SITR-Calib47.50 ± 0.6332.91 ± 1.38N/AN/A
AnyTouch Match45.91 ± 0.5234.86 ± 0.3256.84 ± 1.0749.44 ± 1.25
CTSRL CSM44.91 ± 0.7836.92 ± 0.5459.41 ± 2.2848.98 ± 1.62
BIDETA85.71 ± 0.1287.09 ± 0.6179.13 ± 1.2083.58 ± 1.05

Results average six target sensors for SITR and five for TacVerse. TacVerse lacks the calibration images required to run SITR-Calib.We will supplement Appendix A.4 with these architectures, training configurations, and modifications to the original mechanisms.

6. Are the reported standard deviations incorrect?

We will correct Figure 1 to “TacTip” and “ViTacTip.”

Both results are correct. On Mini_3 internal validation with 5% labels, TVL achieves 92.38 ± 1.91 at 5 iterations and 92.38 ± 2.08 at 10. The same three seeds produce different per-seed results, giving equal means but different standard deviations. The main experiments use 5 iterations. We will move Table 29’s TVL selection marker from 10 to 5, retaining both rows. Main results remain unchanged.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Rebuttal Summary

We thank the area chairs and reviewers for their time and feedback. The discussion helped us clarify BIDETA and its scope, and the additional comparisons strengthened the evidence for label-efficient adaptation to unseen tactile sensors. 

Reviewer ozxj (Rating 3) 

Thank you for your detailed suggestions and for raising your rating to 3 after reviewing the new results. 

  1. Baseline comparisons and adaptation cost. BIDETA outperforms linear classifiers, concatenated shrinkage LDA, and an MLP using the same frozen dual readouts, as well as last-block fine-tuning under the same label budget. We also reported adaptation and query inference times from cached representations and added a frozen DINOv2 evaluation.
  2. Spectral correction and propagation. On SITR with Sparsh and 10% labeled data, conventional label propagation achieves 81.85% and BIDETA achieves 87.09% with LDA predictions and prior normalization held fixed. We will specify the settings for each Table 4 variant and report the validation-selected constant propagation coefficient and its accuracy as mean ± standard deviation, separately from the mean-matched control at 86.12%.
  3. Class priors and task scope. In the same SITR setting, BIDETA without prior normalization still outperforms dual-readout LDA by 4.21 percentage points. We also evaluated long-tailed queries and smaller batches, and tested the variant without prior normalization on batches missing some supported classes. We will clarify the scope of prior normalization under class imbalance and explain that TacQuad evaluates closed-set ranking of enrolled identities on unseen sensors.
  4. Original and supplementary configurations. We clarified that hyperparameters were selected through internal training-set validation and that a disagreement weight of 1 disables the entropy term. We will document the actual disagreement weights and iteration counts used for Tables 1, 2, and 4, and identify the supplementary configurations separately in Table 7.
  5. Baseline implementation. BIDETA retains its advantage when SITR-Calib, AnyTouch Match, and CTSRL CSM use the same dual readouts. We will add the matched-readout results to Table 6 and document the adaptation and training details in Appendix A.4.
  6. Figure and table corrections. We will standardize the TacTip and ViTacTip names in Figure 1 and move the TVL selection marker in Table 29 from 10 to 5 iterations at the 5% label budget, retaining both verified results.  Reviewer 1iEn (Rating 3) 

Thank you for recognizing the practical value of the work and confirming that the matched-readout experiments addressed your baseline concerns. 

  1. Readout selection. Some baselines improve with dual readouts, and BIDETA retains its advantage. We will document the original readout choices and how the baselines were adapted.
  2. Query-specific gating. On SITR with Sparsh and 10% labels, the mean-matched fixed weight yields 86.12% accuracy, and query-specific gating reaches 87.09%. We will include this comparison and explain how the fixed weight was determined.
  3. Class imbalance and smaller batches. Prior normalization gives a small average gain with an approximately 3-to-1 support class imbalance and balanced queries. At query batch sizes of 64 and 256, it improves accuracy by 2.18 and 2.66 percentage points, respectively.
  4. Graph propagation with very few labels. On TacVerse Shape with Sparsh and three support examples per class, the difficulty is already visible in the initial predictions. Adding graph propagation improves average accuracy across three seeds by 0.56 percentage points on TacTip and 0.22 on ViTac.

Reviewer 7dC8 (Rating 1) 

Your questions helped us explain BIDETA’s computational steps, readout definitions, and experimental evidence more clearly. Thank you for engaging in the discussion and confirming that the clarification resolved your concern about the readout definitions. 


We will incorporate these clarifications, corrections, and new results into the revised manuscript. The anonymous code repository is available for review. We will include the repository link in the revised manuscript and publicly release the code alongside the revision to support independent reproduction.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.