acceptodds
Under review as a conference paper at ICLR 2027

Consensus-Anchored Residual Learning for Incomplete Multi-View Multi-Label Classification

Abstract

Incomplete multi-view multi-label classification seeks to leverage heterogeneous information from partially observed views and labels. Existing approaches often couple consensus learning with the extraction of view-specific information, allowing complementary updates to perturb the shared predictor. Under sparse supervision, such coupling may compromise reliable consensus and diminish the predictive gains from view-specific information. To address this issue, we propose Consensus-Anchored Residual Learning (CARE), a new two-stage framework that preserves shared knowledge while learning complementary corrections. Specifically, the first stage establishes a consensus predictor through cross-view fusion guided by label semantics. The second stage anchors this predictor and learns residual refinements around it, while retaining the original consensus as an incumbent during model selection. This design provides a stable reference for exploiting view-specific information and evaluating its predictive contribution. Furthermore, our analysis characterizes conditions for beneficial residual updates and relates the gains retained after selection to conditional complementary information, approximation error, and selection uncertainty. Extensive experiments on six benchmark datasets demonstrate the superiority of CARE over state-of-the-art methods, together with its effectiveness in chest radiograph screening.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Multi-view learning, Multi-label classification, Missing data, Representation learning, Residual learning

Reviews

Among:
Shape:
1234this paper, Top 4%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.33 over 3 reviews, above 96% of the others.
Reviewer_waeHOpenReview ↗
Rating
4
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

CARE addresses incomplete multi-view multi-label classification in two stages. It learns a consensus predictor, freezes its full state, and learns gated residual corrections. Model selection retains the original predictor as a candidate. The paper analyzes when corrections can improve prediction and reports results on six benchmarks, adaptations of four existing methods, and a chest radiograph application.

Critical strengths
  1. CARE preserves a complete consensus predictor while learning residual corrections and keeps it available during selection. This gives complementary learning a stable reference. Improvements when adapting four existing methods suggest that the framework has value beyond its own architecture.
  2. CARE achieves the highest reported AP and AUC on all six benchmarks. Paired-stage results show consistent gains from refinement, while component ablations and missingness studies provide further support.
Critical weaknesses
  1. The synthetic experiment shows that both freezing and continued agreement regularization preserve shared structure. It remains unclear whether freezing offers additional benefits on real datasets.
  2. The auxiliary losses and use of validation data need clearer explanations.
Questions for authors
  1. How does freezing compare with joint refinement that retains the agreement term on real datasets?
  2. Could you provide more details on the training and validation procedures?
1 comment

We thank Reviewer waeH for recognizing the value of a stable consensus reference and CARE’s applicability to existing methods. We address the two concerns below.

W1&Q1: Freezing versus continued agreement on real datasets.

We conducted a matched comparison on Corel5k, OBJECT, and Pascal07 at PER=LMR=50%. For each dataset, all variants start from the same consensus checkpoint selected over 50 Stage-1 epochs and receive 25 refinement epochs. They share the data split, observation masks, residual initialization, random states, optimizer settings, and validation criterion.

Joint refinement updates the shared and residual streams together, including normalization statistics. Joint + agreement additionally retains the original cross-view agreement loss with its original weight. CARE freezes the parameters and normalization state defining the consensus predictor. All variants retain the Stage-1 predictor as a selection candidate.

MethodCorel5k AP / AUCOBJECT AP / AUCPascal07 AP / AUC
Consensus0.4107 / 0.90890.6629 / 0.90720.5622 / 0.8581
Joint0.4237 / 0.90540.6629 / 0.90720.5622 / 0.8581
Joint + agreement0.4274 / 0.90400.6629 / 0.90720.5622 / 0.8581
CARE0.4421 / 0.92170.6654 / 0.90820.5809 / 0.8694

The table reports single-run test scores after validation-based selection. CARE achieves higher AP and AUC on all three datasets. Its AP gains over Joint + agreement are 1.47, 0.25, and 1.87 percentage points on Corel5k, OBJECT, and Pascal07, respectively.

Validation-based selection retains the Stage-1 predictor for both joint variants on OBJECT and Pascal07. The table below reports the test performance of their best refinement checkpoints.

DatasetRefinementCandidate AP / AUCConsensus retained
Corel5kJoint0.4237 / 0.9054No
Corel5kJoint + agreement0.4274 / 0.9040No
OBJECTJoint0.6198 / 0.8767Yes
OBJECTJoint + agreement0.6452 / 0.8917Yes
Pascal07Joint0.5479 / 0.8356Yes
Pascal07Joint + agreement0.5545 / 0.8502Yes

Continued agreement improves the refinement candidate’s AP over unregularized joint optimization on all three datasets. Freezing provides a further benefit in this matched comparison. Agreement encourages consistency among evolving view representations. Freezing additionally preserves the learned prediction function, providing a stable reference for residual correction throughout refinement.

W2&Q2: Auxiliary objectives and the training and validation protocol.

Stage 1 supervises the fused and per-view shared predictions using observed labels. Label-semantic alignment shapes the shared representation, cross-view agreement encourages consistency between available view posteriors and their fused posterior, and reconstruction preserves information from observed inputs.

Stage 2 supervises both the final gated prediction and the per-view corrected predictions. Each per-view auxiliary prediction adds its residual directly to the fixed consensus logit, supplying a learning signal even when its gate is small. Reconstruction and shared/private separation guide complementary representation learning. Gate regularization controls the strength of residual contributions. The Joint + agreement control adds the original agreement term to this Stage-2 objective.

The label mask excludes hidden annotations from classification and semantic alignment. Unavailable view inputs are zeroed before encoding, and view availability is incorporated into fusion, reconstruction, agreement, and separation. Per-view auxiliary classification uses the label mask in Stage 1 and both label and view masks in Stage 2. Feature normalization is fitted separately for each view using its observed training rows.

We use the supplied disjoint training, validation, and test indices. Both stages select checkpoints by validation AP+AUC. A refinement checkpoint replaces the Stage-1 predictor only when its validation score is strictly higher. Macro-F1 thresholds are calibrated on validation predictions. Test evaluation follows completion of training and model selection. All training hyperparameters remain fixed across these comparisons.

Reviewer_wV3gOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

CARE is a two-stage framework for incomplete multi-view multi-label classification. It first learns a consensus predictor and then freezes it while learning gated residual corrections. The original predictor remains available during selection. The paper evaluates its own model on six benchmarks and applies the framework to SIP, MTD, DRLS, and RANK on Corel5k and Pascal07.

Critical strengths
  1. The main strength is its value as a reusable framework. Figure 1(c) reports AP improvements for four existing methods on both tested datasets. This makes the contribution more useful than an improvement to one architecture.

  2. The framework provides a practical way to extend an existing predictor while keeping the original solution available. Its own model also achieves the highest reported AP and AUC on six benchmarks, with further support from stage-wise comparisons and missingness experiments.

Critical weaknesses
  1. The evidence for transferability is narrower than the main benchmark study. The adaptations are tested on only two datasets, and Figure 1(c) does not report variation across runs. This limits confidence in how broadly and consistently the framework helps existing methods.

  2. The adaptation procedure needs a clearer description. It is not clear how much each original method must change or whether the gains persist against simply training it longer. These details matter for judging the framework’s practical value.

Questions for authors
  1. How consistent are the adaptation gains across runs and missingness settings?

  2. What changes are needed to adapt each method, and how do the adapted versions compare with the originals under the same total training budget?

  3. Can the harmful consensus drift described in Eq. (5) be measured on real datasets? Does it explain which methods benefit most from CARE?

  4. How much of the improvement comes from recovering additional view-specific information, and how much comes from correcting errors already recoverable from the consensus features?

  5. When is the additive residual form in Eq. (6) sufficient? Can it recover complementary signals that require interactions between views?

2 comments

We thank the reviewer for recognizing CARE's value as a reusable framework. Additional experiments clarify its transferability and refinement mechanism.

W1 & Q1: Consistency at 30% and 50% missingness.

We extended the transfer study to OBJECT using source-guided implementations of SIP, MTD, DRLS and RANK. Three optimization seeds share fixed sample indices and one observation mask per rate. Paired gains are mean±sample SD in percentage points. CT and FT denote continued training and fine-tuning.

PER=LMRMethodΔAP vs CTΔAP vs FTΔAUC vs CTΔAUC vs FT
30%SIP+0.70±0.65+0.77±0.31+0.36±0.17+0.33±0.20
30%MTD+0.47±0.62+0.47±0.62+0.16±0.22+0.16±0.22
30%DRLS+0.49±0.15+0.49±0.15+0.36±0.08+0.36±0.08
30%RANK+0.80±0.12+1.35±0.85+0.08±0.10+0.28±0.45
50%SIP+0.53±0.46+0.45±0.51+0.29±0.25+0.51±0.44
50%MTD+1.25±0.17+1.29±0.16+0.67±0.04+0.60±0.14
50%DRLS+0.54±0.42+0.54±0.42+0.36±0.06+0.36±0.06
50%RANK+0.35±0.28+0.48±0.44+0.12±0.13-0.01±0.12

All four adaptations improve mean AP over CT and FT at both missingness rates. AP increases in 22 of 24 paired runs against each control, extending the evidence in Figure 1(c) to OBJECT at 30% and 50%.

W2 & Q2: Adaptation procedure and training budget.

SIP and RANK reuse the original prediction path and add gated view-specific residuals. MTD and DRLS learn a separate shared-only reference within their shared/private architecture before residual learning. All adaptations freeze the reference parameters and buffers and retain it during model selection. This gives complementary learning a fixed predictive reference and preserves the learned consensus throughout refinement.

CARE and both controls receive 50+25 supervised epochs. CT continues the original objective and optimizer, whereas FT uses the best original checkpoint with CARE's refinement learning rate, scheduler and ranking loss. DRLS retains native 100-epoch pretraining in all paths. Validation AP+AUC selects every model. The AP gains over both controls support the benefits of two-stage adaptation with matched supervised training epochs.

Q3: Consensus drift and the benefit of anchoring.

We measured Eq. (5) in joint refinement with agreement regularization at the best validation-ranked residual checkpoint, before comparison with the preserved consensus. Holding its correction dd fixed, we restored the original consensus and computed Δdrift=R(F∗+d+e)−R(F∗+d)\Delta_{\mathrm{drift}}=R(F_*+d+e)-R(F_*+d), where ee is consensus-logit drift and RR is mean logistic loss.

DatasetJoint + agreementConsensus restoredRelative reduction
Corel5k0.046980.046121.83%
OBJECT0.167760.1460712.93%
Pascal070.244650.1976119.23%

Restoring the consensus lowers loss on all three datasets, revealing a measurable drift cost under agreement regularization. CARE removes this term by fixing the complete consensus state, so e=0e=0. Our analysis relates the attainable gains to the complementary information available to the residual and how closely it can approximate the required correction.

Q4: Contributions of consensus correction and view access.

We trained a residual using only the frozen consensus mean, log variance and availability mask. Downstream layers, losses, the 25-epoch schedule and selection were unchanged, with input-layer sizes adjusted to the available inputs. Below is single-run test AP at PER=LMR=50% on the original partitions.

DatasetConsensusConsensus-only residualCARE
Corel5k0.410670.413790.44214
OBJECT0.662870.663250.66539
Pascal070.562160.572300.58090

The consensus-only residual improves AP on all three datasets. View access adds approximately 2.84, 0.21 and 0.86 percentage points on Corel5k, OBJECT and Pascal07. The preserved consensus thus provides a common reference for assessing improvements from better use of existing representations and from direct view access.

Q5: Additive corrections and cross-view interactions.

CARE preserves the cross-view interactions learned by the consensus while refining its predictions through additive corrections. Equation (6) is sufficient when the remaining conditional log-odds error is representable by the gated per-view residuals across observation masks. For remaining interactions requiring joint access to multiple views, a joint-view residual would expand the correction family while retaining the same anchoring principle. The residual approximation term describes how closely this family can represent the required correction.

I thank the authors for the thorough rebuttal. The extended transfer study, matched-budget controls, and consensus-drift measurements address my concerns. My score remains unchanged.

Reviewer_gcR5OpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes a new two-stage framework called Consensus-Anchored Residual Learning (CARE) that preserves shared knowledge while learning complementary corrections for incomplete multi-view multi-label classification tasks. The first stage establishes a consensus predictor through cross-view fusion guided by label semantics. The second stage anchors this predictor and learns residual refinements around it, while retaining the original consensus as an incumbent during model selection. Extensive experiments on six benchmark datasets demonstrate the superiority of CARE over state-of-the-art methods. This paper is well written, and the overall readability is good. While this paper is above average in quality, it still has some shortcomings that need to be addressed.

Critical strengths
  1. The research motivation is clear, and the issue identified in the paper that complementary updates to perturb the shared predictor during joint optimization exactly exists.
  2. The proposed method is straightforward and easy to understand and is supported by an intuitive graphical framework.
  3. The authors conduct extensive experiments, and the results of comparisons with state-of-the-art methods validate the effectiveness of their proposed approach.
Critical weaknesses
  1. While the author's theory may be mathematically sound, it may diverge from the experiments presented in this paper—specifically, the theory requires that the selection of samples be independent of the candidate construction, whereas in the experiments, both the consensus checkpoints and the residual candidates are selected based on the mean AP/AUC scores from the validation set.
  2. The model proposed by the authors contains parameters such as α,β,η,γ, and λgateλ_{gate}; consequently, the hyperparameter tuning and model selection may be challenging. Furthermore, the authors do not provide details of hyperparameter selection, nor do they specify optimal parameter values for α,β,η,γ,λgateλ_{gate}, making the results difficult to reproduce.
  3. Two-stage approaches are not new in the field of incomplete multi-view multi-label learning; beyond the TSIEN method cited by the authors, there are others such as MVFD (WACV 2025). Specifically, MVFD also freezes the representations (the consensus encoder) learned during the first stage. What is the essential difference between the proposed approach and these methods?
Questions for authors
  1. If a significant proportion of views and labels are missing, the prototypes learned in the first stage may be unreliable; how do the authors address this issue?
  2. How does CARE compare to the nine methods in terms of training time and inference overhead? Does the two-stage training process introduce any additional overhead?
  3. Proposition 3.1 provides a decomposition under the BCE risk, whereas the experiments primarily rely on AP/AUC for model selection and evaluation. Is it possible to derive an analogous relationship for the AP/AUC metrics involving residual complementary information and approximation error?
3 comments
AuthorsOpenReview ↗

Checkpoint selection under shared validation and bounds on ranking gains

We thank the reviewer for highlighting the dependence between the two checkpoint choices. Section 3.3 now states the guarantee for a shared validation set, and Appendix A gives the proof.

W1. Checkpoint selection with a shared validation set

Stage 1 selects the anchor on validation. Stage 2 trains residuals from that anchor and retains it as a candidate on the same validation set. The independence condition applies to the family of possible checkpoints. Both selections may depend on validation scores.

Let G1,…,G50G_1,\ldots,G_{50} be the consensus checkpoints. For each possible anchor GjG_j, the training data, fixed update schedule, and training randomness determine 25 residual checkpoints Fj,1,…,Fj,25F_{j,1},\ldots,F_{j,25}. Write Fj,0=GjF_{j,0}=G_j for the archived anchor. Under a fixed configuration, this gives at most 50(25+1)=130050(25+1)=1300 potential predictors specified independently of the validation values. Validation selects the anchor index JJ. Training then follows the prescribed updates for that anchor.

Both stages maximize the validation estimate S^\widehat S of S=(AP+AUC)/2S=(\mathrm{AP}+\mathrm{AUC})/2. Stage 1 selects the best consensus checkpoint GJG_J. Stage 2 selects the best candidate on its branch, including GJG_J. The validation score of the returned predictor F^\widehat F therefore matches or exceeds every consensus score and every score on the selected branch. Whenever ∣S^(F)−S(F)∣≤εS|\widehat S(F)-S(F)|\leq\varepsilon_S holds throughout the potential family,

S(F^)≥max⁡(max⁡jS(Gj),max⁡tS(FJ,t))−2εS.S(\widehat{F}) \geq \max(\max_{j} S(G_{j}), \max_{t} S(F_{J,t})) - 2\varepsilon_{S}.

The bound holds for all anchors simultaneously, including the one selected on validation. The term 2εS2\varepsilon_S accounts for uncertainty in the two scores being compared. Appendix A assumes i.i.d. validation examples independent of training. The same proof covers the union of candidate families if validation also selects among prespecified configurations.

Q3. Extending Proposition 3.1 to AP and AUC

We appreciate the reviewer’s suggestion to extend the risk decomposition to the ranking metrics used for model selection. For population binary cross-entropy, the reduction in optimal loss equals the additional information the residual inputs provide about each label beyond the consensus, averaged over labels. For AP and AUC, the corresponding term measures how much those same inputs improve the best attainable ranking score. Appendix A gives the decomposition for each metric and for their mean.

Let ZZ be the information available to the frozen consensus, including the view mask, and UU the observed inputs available to the residual. For a ranking metric MM, let VM(Z)\mathcal V_M(Z) and VM(Z,U)\mathcal V_M(Z,U) denote the best population scores attainable with each set of inputs, allowing independent randomization. Their difference JM\mathcal J_M measures the ranking value of UU given ZZ. The anchor F∗=GJF_{\ast}=G_J falls short of its optimum by e∗,M=VM(Z)−M(F∗)e_{\ast,M}=\mathcal V_M(Z)-M(F_{\ast}). The corresponding gap for a refinement FkF_k is ek,M=VM(Z,U)−M(Fk)e_{k,M}=\mathcal V_M(Z,U)-M(F_k). Subtracting gives

M(Fk)−M(F∗)=JM+e∗,M−ek,M.M(F_k)-M(F_{\ast})=\mathcal J_M+e_{\ast,M}-e_{k,M}.

As in Proposition 3.1, the first term measures what the additional inputs make possible. The remaining terms capture how well each predictor uses its inputs. A refinement can therefore gain from new information or from correcting ranking errors the anchor made with the information already available to it.

Conditional information bounds JM\mathcal J_M. For instance-averaged AP, JAP≤I(Y;U∣Z)/2\mathcal J_{\mathrm{AP}}\leq\sqrt{I(Y;U\mid Z)/2}. The full label vector YY enters because AP depends on which labels are relevant together. For the benchmark’s pooled top-kk AUC, the bound uses the quantity from Proposition 3.1, Icomp=c−1∑j=1cI(Yj;U∣Z)\mathcal I_{\mathrm{comp}}=c^{-1}\sum_{j=1}^c I(Y_j;U\mid Z). With cc labels and an expected count of μ\mu positive labels per example, where 0<μ<c−10 < \mu < c-1,

0≤JAUC≤c(12μ+1c−1−μ)Icomp2.0\leq\mathcal J_{\mathrm{AUC}}\leq c\left(\frac{1}{2\mu}+\frac{1}{c-1-\mu}\right)\sqrt{\frac{\mathcal I_{\mathrm{comp}}}{2}}.

CARE selects checkpoints by S=(AP+AUC)/2S=(\mathrm{AP}+\mathrm{AUC})/2. Applying the identity to this score and using the uniform validation bound in W1 gives

S(F^)−S(F∗)≥JS+e∗,S−min⁡kek,S−2εS,S(\widehat F)-S(F_{\ast})\geq \mathcal J_S+e_{\ast,S}-\min_k e_{k,S}-2\varepsilon_S,

where kk ranges over checkpoints on the selected branch, including the anchor. This bounds the population gain of the selected predictor in terms of the ranking value of the residual inputs, the approximation errors, and validation uncertainty. On the same event, a validation gain over the anchor exceeding 2εS2\varepsilon_S implies a positive population gain in SS.

AuthorsOpenReview ↗

Training configuration and prototype learning under missing observations

W2. Hyperparameter settings and checkpoint selection

We used the same loss weights and training configuration across all six benchmarks.

SettingValue
Reconstruction weight α\alpha11
Semantic alignment weight β\beta10−310^{-3}
Cross-view coherence weight η\eta11
Disentanglement weight γ\gamma10−210^{-2}
Residual gate penalty λgate\lambda_{\mathrm{gate}}10−210^{-2}
Training epochs50 for consensus, 25 for refinement
Initial learning rate for Adam10−310^{-3} for consensus, 5×10−45\times10^{-4} for refinement
Batch size128
Latent dimension512
Semantic alignment temperature0.1

Figure 7 examines β\beta for consensus alignment and λgate\lambda_{\mathrm{gate}} for residual gating. On OBJECT and Pascal07, neighboring settings within the tested ranges yield similar AP.

Normalization statistics are fitted on observed training rows and reused for validation and testing. Stage 1 retains the latest checkpoint attaining the best validation AP+AUC. Stage 2 starts with this anchor and replaces the current best checkpoint only if the score improves. After checkpoint selection, we choose one Macro-F1 threshold per stage on the validation set and apply it to all labels. AP and AUC are computed from continuous scores. Appendix E specifies the masking and selection rules. We will include the four anonymized core files covering the model, losses, training and evaluation, and data loading.

Q1. Prototype estimation under missing views and labels

Sparse annotations reduce the direct supervision available for individual label prototypes. CARE estimates the prototypes through classification and semantic alignment. The classification loss fits predictions to the observed labels. Semantic alignment contrasts each observed positive label with the observed negative labels of the same example. Reconstruction and cross-view coherence also constrain the shared representation. Gaussian fusion uses the available view experts together with the prior, which keeps the fused distribution well defined as views become unavailable.

The masks also govern auxiliary supervision. Stage 1 supervises per-view predictions from masked inputs at observed label coordinates. Stage 2 supervises each view-specific correction only when both the view and the label are observed. Appendix E specifies how the masks enter each objective.

Stage 2 preserves the prototype’s contribution to the consensus score. If that score is inaccurate, a residual can correct it using the available views. Validation then compares the refined predictor with the original consensus. At 30% view and label missingness, the selected predictors improve mean AP over the anchor on all six benchmarks, with relative gains from 1.08% to 13.41%.

AuthorsOpenReview ↗

Preservation of the consensus predictor and computational cost

W3. Prediction refinement in CARE, TSIEN, and MVFD

TSIEN and MVFD both use a first stage to learn representations for a subsequent classifier. TSIEN passes the hidden features of its view-specific classifiers to a network that aligns and fuses them. MVFD freezes the consistent encoders and trains a second classifier on the element-wise product of the shared features and the gated view-specific features.

CARE preserves the complete first-stage predictor, including its classification head, label prototypes, and normalization state. For each observed input xx, the second stage learns a correction to the original label logits,

Fϕ(x)=F∗(x)+Dϕ(x).F_\phi(x)=F_{\ast}(x)+D_\phi(x).

Setting Dϕ(x)=0D_\phi(x)=0 reproduces the first-stage predictions for every admissible view mask. The same predictor is an explicit candidate during validation, so a refinement is retained only if it improves the selection score over the anchor. The frozen predictor thus serves as the reference for both learning a correction and assessing the refined checkpoint. We have clarified this distinction in the introduction and related work.

The table lists reported mean AP and AUC on the five shared benchmarks at 50% missing views and 50% missing labels. Tan et al. report the TSIEN results in Table 2 of their AAAI 2024 paper, and Xie et al. report the MVFD results in Table 1 of their WACV 2025 paper.

DatasetTSIEN APMVFD APCARE APTSIEN AUCMVFD AUCCARE AUC
Corel5k0.4360.4410.4500.9200.9080.923
ESPGame0.3190.3260.3440.8630.8590.869
IAPRTC120.3610.3600.3710.8990.8960.902
MIRFlickr0.6310.6150.6450.8720.8730.876
Pascal070.5810.5640.5990.8680.8600.875

CARE has higher reported means for both metrics on all five datasets. Our component ablation isolates the effect of preserving the predictor under matched splits and masks. On Pascal07, with gating and incumbent selection enabled in both variants, freezing the consensus raises AP from 0.587 to 0.599 and AUC from 0.863 to 0.875. Figure 1(c) further shows that applying the same framework to SIP, MTD, DRLS, and RANK improves AP on both Corel5k and Pascal07.

Q2. Training time and inference overhead

The table reports training times in seconds for MTD, SIP, and CARE. CARE’s totals include both stages.

DatasetMTDSIPCARE
Pascal071457583727
Mirflickr563727749
ESPGame408116051824
Corel5k252110401423

To measure the cost of a refinement update, we compared frozen and trainable consensus predictors on the same batch of 128 OBJECT examples, using the same architecture, objective, and latent dimension of 512. Freezing the consensus leaves the residual and reconstruction parameters trainable and reduces the number of trainable parameters by 40.2%. Across six interleaved timing blocks, a refinement update took 38.9% less time on average with the consensus frozen.

At inference, CARE evaluates the predictor selected on validation. For a batch of 128 examples, the CPU prediction pass averaged 55.0 ms for the consensus alone and 160.8 ms with the residual enabled.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.