acceptodds
Under review as a conference paper at ICLR 2027

STAGE: Constrained Graph-Phase Refinement for Binary Hashing

Abstract

Learning binary hash codes for image retrieval requires preserving semantic relationships while enabling independent image encoding. However, jointly refined targets can vary with batch composition or exceed the encoder’s output range, and lower refinement energy does not necessarily yield better retrieval. We propose STAGE, a constrained graph-phase framework that learns an independent hash encoder from jointly refined targets. In its minibatch formulation, a graph-phase teacher combines graph smoothness, semantic supervision, phase separation, and bit decorrelation under a zero-mean constraint. For a fixed graph and energy, learnable per-bit positive-definite spectral operators adapt refinement dynamics while preserving the constrained stationary points. To facilitate target transfer, a fixed-reference extension produces bounded, zero-mean targets, caches them by image identity, and trains the teacher using the student’s ranking loss after a virtual update. We establish mean conservation, conditional energy descent, and a bound relating binary imbalance to quantization error. Experiments on CIFAR-10, CIFAR-100, and VOC 2007 using frozen ResNet-18 features and 8-, 16-, and 32-bit codes show that both formulations achieve the highest mean mAP among the compared methods in all nine settings of their respective protocols. Under the fixed-reference protocol, STAGE improves mean mAP over matched scalar refinement by 0.15–3.91 percentage points. Graph computations are confined to training, retaining independent encoding and standard Hamming search at inference.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
Binary Hashing, Image Retrieval, Constrained Graph-Phase Refinement, Learnable Spectral Operators, Teacher-Student Learning, Meta-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_NzrtOpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

STAGE refines continuous codes on a graph and uses them as targets for a binary hash encoder. The energy combines graph structure, quantization and semantic/code regularization. Projected updates preserve zero-mean codes, while learned positive per-bit spectral operators control refinement. The student fits these targets and encodes each sample independently at inference. Alongside batch refinement, the paper introduces a fixed-reference formulation with bounded centered readout, target caching by sample identity, and teacher feedback through a virtual student update. The theoretical results cover conservation, energy descent and binary imbalance under stated conditions. Experiments use frozen ResNet-18 features on CIFAR-10, CIFAR-100 and VOC 2007 with 8, 16 and 32 bits.

Critical strengths
  1. The paper addresses a concrete deployment problem: codes optimized jointly on a graph must be learnable by an encoder that processes samples independently. The fixed-reference construction handles target range, centering and sample identity while retaining ordinary encoding and Hamming retrieval at inference. The theory also states the conditions under which its results hold; it does not claim an unconditional retrieval guarantee.

  2. The internal controls are a strength. Direct-energy, scalar, bounded and optimizer-consistent feedback variants help separate refinement from generic regularization. Meta-scalar is particularly relevant because it shares the readout, refresh and outer supervision, isolating the additional spectral shape. STAGE achieves the highest reported means across the tested settings, so the evidence extends beyond a comparison with the base model.

Critical weaknesses
  1. The small margins need a more reliable estimate of uncertainty. In the nine fixed-reference settings, gains over the strongest control range from 0.12 to 0.71 percentage points. These are the practical gains over available alternatives; the more closely matched Meta-scalar comparison answers the separate question of what spectral shape adds. Updated Table 5 means for CIFAR-100 and VOC lack updated standard deviations. Even marginal means and SDs would not establish uncertainty in the paired differences.

  2. It is unclear whether the extra training machinery earns its cost over simpler alternatives. At 8 bits on CIFAR-10, Bounded T=0 reaches 69.78 mAP compared with 70.27 for the full model. At 8 bits on VOC, the gain over Reference base is only 0.54 points, despite a larger margin over Meta-scalar. Fixed-reference training takes about 2.31-2.39 times as long as the base model and about 1.10 times as long as Meta-scalar. The equal-time experiment covers only a subset of the batch protocol, leaving the fixed-reference results without that comparison.

Questions for authors

Which conclusions depend on the VOC data accessed during the pilot? Can you verify them with frozen hyperparameters on a split or setting that was not used in development? Please distinguish these results from the independent CIFAR validation/final-query protocol.

4 comments
AuthorsOpenReview ↗

Response to W[1/3]

Thank you for distinguishing the contribution of spectral shape from the practical gains over competitive alternatives, and for recognizing the value of the matched controls. We agree that the small margins make uncertainty in the paired differences important for interpreting these results.

Meta-scalar shares the bounded readout, target refresh, and outer supervision with STAGE-Ref, providing a matched comparison for assessing the contribution of spectral parameterization. Bounded T=0 provides a simpler target construction for assessing the benefit of the full refinement procedure.

We compute each gain from the displayed mean mAP values.

ΔmAP=mAP‾STAGE-Ref−mAP‾comparator.\Delta \mathrm{mAP} = \overline{\mathrm{mAP}}_{\text{STAGE-Ref}} - \overline{\mathrm{mAP}}_{\text{comparator}}.

Table R1. Eight-bit fixed-reference retrieval results from Tables 5–6. mAP is summarized over three seeds, with sample SDs shown where available. ΔmAP is in percentage points (pp).

DatasetComparatorComparator mAP (%)STAGE-Ref mAP (%)ΔmAP (pp)
CIFAR-10Meta-scalar69.89 ± 0.4970.27 ± 0.22+0.38
CIFAR-10Bounded T=069.78 ± 0.4370.27 ± 0.22+0.49
CIFAR-100Meta-scalar9.18 ± 0.329.98†+0.80
CIFAR-100Bounded T=09.70 ± 0.839.98†+0.28
VOC 2007Meta-scalar64.44 ± 0.7568.35†+3.91
VOC 2007Bounded T=064.97 ± 0.2968.35†+3.38

† The sample SD corresponding to the updated mean is unavailable.

At 8 bits, the mean gains over Meta-scalar are 0.38, 0.80, and 3.91 pp on CIFAR-10, CIFAR-100, and VOC 2007, respectively. Across the nine settings in Table 2, the gains over the strongest control evaluated in each setting range from 0.12 to 0.71 pp.

The reported sample SDs describe seed-to-seed variation within each method. Quantifying uncertainty in the paired differences requires matched seed-level scores, including those underlying the updated CIFAR-100 and VOC 2007 means. We therefore interpret the results as descriptive evidence of higher mean mAP under the reported settings. Statistical significance and consistency across seeds remain to be established.

AuthorsOpenReview ↗

Response to W[2/3]

Thank you for highlighting the practical trade-off between retrieval performance and training cost. Your examples show why this should be assessed against simpler alternatives as well as the closely matched Meta-scalar control.

We summarize the existing performance and timing results below to make these comparisons explicit.

Table R2. Eight-bit fixed-reference retrieval performance and training time from Tables 5 and 10(b). Results are three-seed means, with sample SD where available. Training time covers 60 epochs and excludes feature extraction, graph/pair/triplet setup, validation, and retrieval.

DatasetMethodmAP (%)Training time (s)
CIFAR-10Reference base66.94 ± 4.457.50 ± 0.17
CIFAR-10Bounded T=069.78 ± 0.439.82 ± 0.41
CIFAR-10Meta-scalar69.89 ± 0.4915.91 ± 0.52
CIFAR-10STAGE-Ref70.27 ± 0.2217.44 ± 0.08
VOC 2007Reference base67.81 ± 0.135.37 ± 0.21
VOC 2007Bounded T=064.97 ± 0.296.39 ± 0.44
VOC 2007Meta-scalar64.44 ± 0.7511.27 ± 0.23
VOC 2007STAGE-Ref68.35†12.42 ± 0.63

† The sample SD corresponding to the updated VOC mAP mean is unavailable.

On CIFAR-10, STAGE-Ref improves over Bounded T=0 by 0.49 percentage points while requiring approximately 1.78 times its training time. On VOC, the improvement over Reference base is 0.54 points at approximately 2.31 times the training time. We agree that these are modest accuracy gains with a measurable computational cost.

The comparison with Meta-scalar addresses the incremental cost within the shared fixed-reference training design. STAGE-Ref requires approximately 10% more training time than Meta-scalar on these two datasets. However, the larger VOC gain over Meta-scalar should be considered alongside the smaller gain over Reference base when assessing practical value.

The additional graph refinement and teacher computations are confined to training. At inference, the method retains independent image encoding and standard Hamming-distance retrieval. This preserves the deployment procedure, while the additional training cost remains a relevant consideration.

We therefore describe the reported improvement as an accuracy–training-cost trade-off under the fixed-epoch protocol. The existing equal-time results in Table 9 concern the batch protocol. An equal-time advantage for STAGE-Ref remains unverified, and we do not extend the batch-protocol evidence to that claim.

AuthorsOpenReview ↗

Response to W[3/3]

Thank you for asking which conclusions depend on the VOC data accessed during the pilot. This is an important distinction between separation within the final evaluation protocol and independence from earlier development decisions.

Appendix A.3.1 discloses that the official VOC test set was accessed during the initial pilot and subsequently reused for final evaluation. The CIFAR final-query blocks were unused in that pilot. We summarize the relevant protocol details below.

Table R4. Validation and final-query usage in the reported evaluation.

DatasetValidation queriesFinal queriesFinal-query block accessed during the initial pilot
CIFAR-105001,000No
CIFAR-1005001,000No
VOC 20075004,952Yes

For VOC, the validation queries are drawn from trainval, and all 4,952 official test images serve as final queries. The final queries are excluded from the retrieval database. These protocol details establish the separation used in the final runs, but the extent to which the earlier VOC observations influenced method design or experimental choices remains undocumented.

Consequently, all reported VOC retrieval comparisons inherit this limitation. This includes the 8-bit gains of 3.91 percentage points over Meta-scalar and 0.54 points over Reference base, as well as broader empirical claims that rely on the VOC results. We therefore treat these measurements as results on a benchmark accessed during development, with the corresponding qualification.

The CIFAR results use their separately reserved final-query blocks and support the reported conclusions on those datasets. They do not independently verify the specific VOC findings. The mathematical results on conservation, stationary points, and conditional descent are derived under their stated assumptions and do not depend on VOC measurements.

We agree that a stronger confirmation would freeze the method, hyperparameters, and checkpoint-selection rule before evaluation on a split or setting genuinely unused during development. Fixing hyperparameters and rerunning the previously accessed VOC test set would not restore that independence. Such an independent confirmation remains outstanding in the current evidence.

Our response therefore makes the provenance and scope of the VOC findings explicit and keeps them separate from the CIFAR final-query protocol. We appreciate your careful attention to this distinction, together with your assessment of the independent-encoding formulation and matched controls. Your comments help us state both the contribution and the remaining evidential limitations more precisely.

AuthorsOpenReview ↗

Official Comment by Authors

We sincerely thank you for your careful review and your recognition of the constrained formulation and independent encoding at inference. Your questions about the small spectral gains, their uncertainty, and the relationship between energy reduction and retrieval performance highlight where the evidence needs a more precise interpretation. Your attention to computational cost, evaluation scope, and the VOC pilot also helps clarify the conditions under which the results should be understood. We appreciate the rigor of your assessment and hope that our point-by-point responses make the contribution, supporting evidence, and remaining limitations clearer. We would welcome any further questions.

Reviewer_5MFwOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes STAGE, a constrained graph-phase refinement framework for binary hashing. It formulates hash-target refinement as a graph-based energy optimization problem that integrates graph smoothness, semantic supervision, a double-well potential, and bit decorrelation under a zero-mean constraint. Learnable per-bit spectral operators are introduced to adapt the refinement dynamics while preserving constrained stationary points under fixed graph and energy conditions. The paper further presents STAGE-Ref, a fixed-reference extension that incorporates student-ranking feedback to guide target refinement. Theoretical analysis examines mean conservation, conditional energy descent, and the relationship between quantization error and binary imbalance. Experiments evaluate the proposed methods against existing hashing approaches.

Critical strengths

1.The paper provides a constrained optimization formulation for graph-based hash target refinement and analyzes its properties, including mean conservation, conditional energy descent, and the relationship between quantization error and binary imbalance. 2.The evaluation considers both retrieval performance and computational cost. It includes three datasets, three code lengths, ablation studies, and comparisons between spectral refinement and scalar refinement. The proposed framework also retains independent image encoding and standard Hamming-distance retrieval at inference.

Critical weaknesses

1.The current experiments do not clearly isolate the contribution of each component. In particular, the gains from learned spectral refinement over scalar refinement are limited in several fixed-reference comparisons. The theoretical analysis of stationary-point preservation and conditional energy descent does not, by itself, establish that the proposed refinement improves retrieval performance. 2.Experiments are limited to frozen ResNet-18 features on CIFAR-10, CIFAR-100, and VOC 2007. Evaluation with a broader range of recent hashing baselines and end-to-end feature learning would help establish the generality of the results. In addition, the authors disclose that the VOC test set was accessed during an initial pilot. Although the final evaluation protocol is described separately, it remains unclear whether the pilot results influenced method design or experimental choices. Some updated fixed-reference comparisons also lack seed-level results or uncertainty estimates, limiting assessment of the reliability of the reported gains.

Questions for authors

1.Under the fixed-reference protocol, STAGE exceeds meta-scalar by only about 0.38 and 0.80 percentage points on CIFAR-10 and CIFAR-100, respectively. Can the authors provide a seed-paired comparison between spectral and scalar refinement under the same training budget, readout, and outer supervision, together with confidence intervals? If the independent contribution of spectral shape is small, what do the authors consider its main value to be? 2.The paper proves stationary-point preservation and conditional energy descent but explicitly does not guarantee retrieval accuracy. Can the authors provide experimental evidence linking the magnitude of energy decrease or spectral mode decay to final mAP? Are there cases where energy is lower but mAP is worse? 3.The current evaluation uses only frozen ResNet-18 features and three datasets. Can the authors add at least one larger-scale benchmark (e.g., an ImageNet subset or MS-COCO) or a stronger hashing baseline? Does the batch-protocol advantage of STAGE persist under stronger end-to-end baselines? 4.The batch protocol takes about 7.3–7.6 times longer per update than the base model, and fixed-reference training takes about 2.31–2.39 times longer over 60 epochs. How do dense affinity construction and teacher updates scale in memory and time on larger databases? Are sparse-graph or mini-batch sampling alternatives available?

5 comments
AuthorsOpenReview ↗

Response to W[1/2] and Q[1/4]

Thank you for asking how much spectral shaping contributes. STAGE-Ref and Meta-scalar use the same 60-epoch training schedule, bounded readout, cache refresh schedule, and outer ranking supervision. For each seed, the two methods also share the student initialization and batch order. Equation (7) normalizes the average spectral response on the mean-zero subspace before applying the learned per-bit gain. Meta-scalar uses the same gain parameterization with a frequency-independent response. This matched comparison tests whether allowing the response to vary by graph frequency improves retrieval at the same epoch count.

Table R1. Eight-bit fixed-reference retrieval from Table 5. Values are means over three seeds, with sample standard deviations where available.

DatasetMeta-scalar mAP (%)STAGE-Ref mAP (%)Mean gain (pp)
CIFAR-1069.89 ± 0.4970.27 ± 0.22+0.38
CIFAR-1009.18 ± 0.329.98+0.80

We agree that these mean gains should be reported with paired uncertainty. Let dsd_s be the difference in mAP between STAGE-Ref and Meta-scalar for seed ss. Computing the exact paired interval requires all three seed-level differences. The CIFAR-10 summaries do, however, bound its half-width from below. The sample standard deviation of the paired differences satisfies sd≥∣sRef−sscalar∣s_d \geq \lvert s_{\mathrm{Ref}} - s_{\mathrm{scalar}} \rvert, where the two terms on the right are the sample standard deviations for the individual methods. Using the reported standard deviations, the half-width of a conventional paired 95% tt interval with three seeds is bounded below as follows, in percentage points.

t0.975,2sd3≥4.30265∣0.22−0.49∣3≈0.67t_{0.975,2}\frac{s_d}{\sqrt{3}} \geq 4.30265\frac{\lvert 0.22 - 0.49 \rvert}{\sqrt{3}} \approx 0.67.

Even the narrowest interval compatible with the reported CIFAR-10 summaries would include zero. For CIFAR-100 and VOC 2007, we need the corresponding seed-level pairs to compute paired intervals.

Spectral shaping lets the refinement updates vary by graph frequency within a finite trajectory while preserving the constrained stationary set under the conditions in §4. We interpret the modest retrieval gains in the matched comparison with Meta-scalar as descriptive.

AuthorsOpenReview ↗

Response to Q[2/4]

On CIFAR-10 with eight-bit codes, Meta-isotropic has a larger relative decrease in teacher energy than STAGE-Ref (25.27% vs. 21.46%) and lower mean retrieval mAP (67.72% vs. 70.27%).

We report the relative energy decrease in percent as 100(1−R)100(1-R), where RR is the ratio of terminal to starting teacher energy.

R=EA(U(T))EA(U(0))R = \frac{E_A(U^{(T)})}{E_A(U^{(0)})}.

Table R2. Eight-bit CIFAR-10 fixed-reference results from Tables 5 and 8. Energy and agreement metrics are means over three seeds. Hamming mAP is reported as mean ± sample standard deviation over three seeds. Target agreement is evaluated using targets regenerated at the selected checkpoints.

MethodET/E0E_T/E_0Target MSESign error (%)Hamming mAP (%)
Bounded T=0T=01.00000.003950.03469.78 ± 0.43
Energy-spectral0.77380.007110.13170.04 ± 0.56
Meta-isotropic0.74730.012020.37067.72 ± 3.61
Meta-scalar0.78090.007020.14769.89 ± 0.49
STAGE-Ref0.78540.006020.11970.27 ± 0.22

Each ratio uses the starting energy at the corresponding checkpoint as its denominator. A comparison of absolute terminal energies from a common starting state remains open, as does the seed-level relationship between energy reduction and mAP.

This observation motivates the feedback objective in STAGE-Ref. The method learns the spectral parameters using the student’s ranking loss after a virtual student update, giving the teacher feedback on how its targets affect student ranking. The stationary-point and descent results characterize the refinement updates, and Figure 3 illustrates conditional modal decay rates on constructed quadratic objectives. Whether modal decay predicts retrieval performance on real image tasks remains open.

AuthorsOpenReview ↗

Response to W[2/2] and Q[3/4]

We agree that the frozen-feature protocol limits the scope of the Batch comparison. All reported image experiments use frozen features from a ResNet-18 pretrained on ImageNet-1K and a shared hash-head architecture. Table 1 compares STAGE-Batch with MDS, SHC, CRH, and CSQ adapted to this setup.

Table R3. STAGE-Batch mAP gains over the strongest published baseline listed for each setting in Table 1. Gains are in percentage points. The baseline is shown in parentheses.

DatasetDatabase size8 bits16 bits32 bits
CIFAR-1050,000+0.48 (MDS)+0.93 (SHC)+0.63 (SHC)
CIFAR-10050,000+0.07 (SHC)+0.34 (SHC)+0.72 (CRH)
VOC 20075,011+3.30 (CRH)+0.24 (CRH)+0.17 (MDS)

The margins vary across datasets and code lengths, and several are below half a percentage point. Whether the gains extend to a larger benchmark or persist in a matched comparison with stronger end-to-end baselines remains open.

VOC development access

As disclosed in Appendix A.3.1, the initial pilot accessed the official VOC 2007 test set, which also serves as the query set for the reported VOC results. The paper specifies the final training, validation, and query splits. Whether observations from the pilot influenced earlier design or experimental choices remains undocumented. The final CIFAR query blocks were held out of the pilot. Independent confirmation of the VOC results would require fixing the experimental settings before evaluation on data held out from development.

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

The paper proposes STAGE, a teacher–student framework for binary image hashing that refines targets through a graph-phase energy combining semantic supervision, graph smoothness, phase separation, and decorrelation under a zero-mean constraint. Learnable per-bit spectral operators adapt refinement dynamics while preserving constrained stationary points. A fixed-reference extension introduces bounded targets, identity-based caching, and student feedback through a virtual optimization step. The analysis establishes mean conservation, conditional energy descent, and a quantization–imbalance bound. Experiments use frozen ResNet-18 features on CIFAR-10, CIFAR-100, and VOC 2007 with 8-, 16-, and 32-bit codes. Both variants achieve the highest reported mean mAP among their respective comparators, supported by controlled comparisons, target diagnostics, and computational-cost measurements.

Critical strengths
  1. The method decouples constrained target refinement from independent encoding, using bounded readout and identity-based caching to address target-range and consistency issues. All graph computations remain confined to training, retaining independent encoding and standard Hamming retrieval at inference.

  2. Direct-energy, isotropic, and per-bit scalar controls provide useful comparisons, while gain normalization and the matched meta-scalar control help assess the additional value of spectral shaping. Both variants achieve the highest mean mAP across all nine settings within their respective protocols.

Critical weaknesses
  1. The method introduces several manually specified weights but reports only limited hyperparameter selection. It remains unclear whether the improvements are robust to changes in the teacher-energy and target-matching weights.

  2. Fixed-reference training takes approximately 2.3 times as long as the reference base, while improvements over the strongest controls are only 0.12–0.71 mAP percentage points. Limited equal-time comparisons leave the accuracy–compute trade-off insufficiently justified.

  3. The main properties largely follow standard projection and preconditioned-gradient arguments. The spectral-rate interpretation neglects cross-bit coupling and assumes a shared eigenbasis between the Hessian and graph Laplacian, without validating these conditions on real tasks.

Questions for authors

See Weaknesses.

4 comments
AuthorsOpenReview ↗

Response to W[1/3]

Thank you for asking whether the gains depend on the choice of weights. We cross-referenced the retrieval results in Table 1 with the selected configurations in Table 3. In four of the nine Batch settings, STAGE-Batch and Per-bit scalar refinement use the same student learning rate, target-matching weight, and refinement step cap.

Table R1. Batch retrieval at matched configurations. The table includes all four settings that meet the matching criterion. The student learning rate is 0.002 in every row. The mAP values are three-seed means from Table 1.

DatasetBitsMatching weightStep capPer-bit scalar mAP (%)STAGE-Batch mAP (%)Gain (pp)
CIFAR-1082867.6571.73+4.08
CIFAR-10160.50.572.3375.50+3.17
CIFAR-100320.50.518.8524.80+5.95
VOC 2007820.565.3969.39+4.00

With these three hyperparameters matched, STAGE-Batch achieves higher mean mAP in all four settings, with gains of 3.17 to 5.95 percentage points.

The matching weight λam\lambda_{\mathrm{am}} scales the student–target matching term in Eq. (4).

Lstudent=Lbase(H,S)+λamnK∥H−sg⁡(U(T))∥F2\mathcal{L}_{\mathrm{student}} = \mathcal{L}_{\mathrm{base}}(H,S) + \frac{\lambda_{\mathrm{am}}}{nK}\left\lVert H - \operatorname{sg}(U^{(T)}) \right\rVert_F^2.

As described in Appendix A.3.2, the Batch calibration compared matching weights of 0.5 and 2 at two student learning rates, with the refinement step cap fixed at 2. It then evaluated step caps of 0.5 and 8 using the selected learning rate and matching weight. The teacher-energy coefficients in Eq. (1) were held fixed at 0.2, 0.4, 1, and 0.05 for graph smoothness, the double-well term, semantic supervision, and decorrelation, respectively.

For the fixed-reference comparison, STAGE-Ref and Meta-scalar both use a matching weight of 0.5 and the shared settings listed in Appendix A.3.3. We will clarify that these comparisons characterize performance at the reported configurations. How the gains change when the teacher-energy coefficients or the target-matching weight are varied within a setting remains an open question.

AuthorsOpenReview ↗

Response to W[2/3]

We agree that the additional training time matters given the size of the gains. Across the nine fixed-reference settings in Table 2, STAGE-Ref improves mean mAP by 0.12–0.71 percentage points over the strongest listed control in each setting.

Table 10(b) reports training times for eight-bit codes over 60 epochs. Across the three datasets, STAGE-Ref takes 2.31–2.39 times as long as Reference base and about 1.10 times as long as Meta-scalar. Meta-scalar shares the readout, refresh schedule, and outer supervision with STAGE-Ref, so this comparison estimates the additional cost of spectral shaping under those shared settings.

CIFAR-10 with eight-bit codes and 60 training epochs. Values are means ± sample standard deviations over three seeds, from Tables 5 and 10(b). Training times exclude feature extraction, setup, validation, and retrieval.

MethodHamming mAP (%)Training time (s)
Reference base66.94 ± 4.457.50 ± 0.17
Bounded T=0T=069.78 ± 0.439.82 ± 0.41
Meta-scalar69.89 ± 0.4915.91 ± 0.52
STAGE-Ref70.27 ± 0.2217.44 ± 0.08

At 60 epochs on CIFAR-10, STAGE-Ref improves mean mAP by 0.49 percentage points over Bounded T=0T=0 and takes an additional 7.62 seconds on average.

Table 9 reports equal-time results for STAGE-Batch. For the fixed-reference protocol, Appendix A.8 proposes comparing STAGE-Ref with Meta-scalar and Bounded T=0T=0 under a 30-second training budget on CIFAR-10 with eight-bit codes. For each control, we would compute the paired mAP difference for each of the three seeds.

ds=mAPSTAGE-Ref,s−mAPcontrol,sd_s = \mathrm{mAP}_{\text{STAGE-Ref},s} - \mathrm{mAP}_{\text{control},s}.

We would summarize the mean paired difference with a 95% confidence interval.

dˉ±t0.975,2sd3\bar{d} \pm t_{0.975,2}\frac{s_d}{\sqrt{3}},

where sds_d is the sample standard deviation of the three paired differences. The reported fixed-reference results quantify the accuracy–time trade-off at a shared epoch count. Equal-time performance against Reference base remains open.

AuthorsOpenReview ↗

Response to W[3/3]

We agree that mean conservation, stationary-set preservation, and conditional descent follow from standard arguments for projection and positive-definite preconditioning. These properties characterize the refinement updates that generate the student’s targets. In §4, the graph, energy function, and mobility are fixed along each refinement trajectory. Let gk=P∇U:kE(U)g_k = P\nabla_{U_{:k}}E(U) denote the projected gradient of the full coupled energy with respect to bit kk.

Table R3. Theoretical results and their conditions. The first three rows concern the full coupled updates. The final row concerns the fixed-bit linearization in Appendix A.1.3.

ResultMathematical statementRequired conditions
Mean conservation1⊤U(t)=0\mathbf{1}^{\top}U^{(t)} = 0Zero-mean initialization and projected updates.
Constrained stationary-set preservationPMk(L)gk=0  ⟺  gk=0PM_k(L)g_k = 0 \iff g_k = 0The mobility Mk(L)M_k(L) is positive definite and commutes with the projection PP.
Discrete energy descentE(U(t+1))≤E(U(t))E(U^{(t+1)}) \leq E(U^{(t)})The gradient is locally Lipschitz along the update segment, and the effective step size satisfies the bound in Appendix A.1.2.
Local modal decaye˙j=−mk(λj)hjej\dot{e}_j = -m_k(\lambda_j)h_j e_jThe linearization is taken at a constrained stationary point with cross-bit Hessian blocks omitted. On the mean-zero subspace, the fixed-bit Hessian is positive definite and shares an orthonormal eigenbasis with the graph Laplacian.

Under the conditions in the final row, mode jj has the continuous-time solution

ej(t)=ej(0)exp⁡[−mk(λj)hjt]e_j(t) = e_j(0)\exp\bigl[-m_k(\lambda_j)h_j t\bigr],

where λj\lambda_j is the corresponding graph Laplacian eigenvalue and hj>0h_j > 0 is the Hessian curvature along that mode. Figure 3 illustrates this local model on constructed quadratic objectives. The image-task diagnostic in Appendix A.6.1 measures initial update directions. Whether the image-task Hessians approximately satisfy the assumptions of the local model remains an open question.

For target generation, the relevant question is how mobility affects the targets reached within a finite refinement budget. Two mobilities that preserve the same stationary set can produce different targets after the same number of steps. STAGE-Ref learns its spectral parameters using the ranking loss evaluated after a virtual student update. The comparison with Meta-scalar evaluates the effect of spectral shaping on retrieval while holding the readout, refresh schedule, and outer supervision fixed. We have clarified in §4 the conditions under which each theoretical result applies.

AuthorsOpenReview ↗

Official Comment by Authors

We sincerely thank you for your careful and constructive review and for recognizing the value of our controlled comparisons. We particularly appreciate the rigor with which you examined how to interpret the reported gains. Your questions about the teacher-energy and target-matching weights prompted us to clarify what the matched-configuration comparisons establish about the role of those weights. Given the modest mAP gains of STAGE-Ref, your request for equal-time comparisons highlighted the importance of evaluating performance under a common training budget. Your observations about cross-bit coupling and the shared eigenbasis also helped us state the assumptions behind the local spectral-rate analysis more explicitly and clarify its relevance to the image retrieval experiments. We believe your feedback has made the paper more precise about what the current results support and which questions remain open. We would be grateful for any further questions or suggestions.

Comments on the paper

1 comment

Dear Reviewers,

The authors’ rebuttal is now available. Please go through it carefully, consider whether it resolves the concerns you raised, and follow up with a comment to the authors.

Thank you for your time and constructive feedback.

AC

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.