acceptodds
Under review as a conference paper at ICLR 2027

Geometric Mean Pooling for Equal-Weight Multiplicative Coarse-Graining

Abstract

As an alternative to the additive and extremal biases of average and max pooling, we introduce Geometric Mean Pooling (GMP), a signed pooling operator that combines the product of feature signs with the geometric mean of feature magnitudes. Motivated by local-to-global composition in quantum many-body physics, GMP retains both joint sign information and a characteristic multiplicative scale without introducing learnable pooling parameters. We show that non-overlapping hierarchical GMP preserves the corresponding global multiplicative statistic and evaluate it on synthetic sequence tasks, iterative coarse-graining, image classification, and molecular lipophilicity regression. On the synthetic tasks, GMP recovers product-based signals more accurately than average and max pooling and maintains predictive performance under the tested levels of multiplicative input noise. On image and molecular data, however, its effectiveness depends on the representation, target parameterization, and placement of local and global pooling. These results position GMP as a complementary, regime-dependent inductive bias for tasks in which equal-weight multiplicative composition is plausible, rather than as a universal replacement for standard pooling operators.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
geometric mean, multiplicative coarse-graining, pooling operators, sign parity, renormalization group

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_Wd85OpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper introduces a pooling operator Geometric Mean Pooling (GMP) computed by combining the product of feature signs with the geometric mean of the magnitudes. The authors compare GMP to max and average pooling operations on synthetic and real datasets including classification and regression tasks. In the synthetic tasks, the labels were generated with signed parity and geometric mean statistics. Accordingly, GMP consistently outperforms average and max pooling in those tasks. The experiments also show GMP keeps the performance advantage under increasing strengths of noise and an ablation without the signed term. On image classification tasks, GMP consistently underperforms or performs similarly to the other operations. On a molecular regression task, using a transformed exponential scale one of the configurations that includes GMP has a slight advantage. On the original scale all configurations with GMP underperform.

Critical strengths

Evaluations on the synthetic datasets. The paper includes extensive evaluations on synthetic datasets, including classification and regression tasks. On these engineered tasks, GMP has a clear advantage over max and average pooling. The authors also present an ablation on the signed term for the classification tasks. Empirical workflow. Multiple seeds in most experiments, evaluation of local and global placement combinations for pooling operations, open code.

Critical weaknesses

Limited methodological novelty. The unsigned geometric mean pool was established in previous work [1], so the novelty of the proposed method here is limited to the introduction of the signed term. Limited comparisons with other pooling operations. GMP is only compared against average and max pooling for the main performance experiments. In particular, comparison against the unsigned geometric mean is critical to establish the advantage of the new method. Weak results on image classification. GMP performs similarly or substantially worse than max/avg pool on all the image classification tasks. Inconclusive results on lipophilicity regression. On the lipophilicity regression task with the transformed task, global GMP and local max have the best mean, but the standard deviation bars overlap with the other operations under local max. It is unclear if the global GMP advantage is statistically significant or not. Furthermore, without transforming the labels, GMP consistently underperforms compared with the pooling alternatives (Appendix E). Lipophilicity experiment reduces to geometric mean pool. In the lipophilicity model, since the convolutional blocks use a softplus activation, the inputs to the pooling operations are all positive. In this case GMP reduces down to the established unsigned geometric mean pool operation, so in any case this experiment does not evaluate the sign term which is the novel component in the paper. Lipophilicity baselines. The motivation behind the lipophilicity baseline comparison is unclear. If the motivation is a comparison with the paper’s 2D CNN then they are not directly comparable as the baselines use the standard scale while the 2D CNN model uses the transformed target. If it is to motivate the use of the Morgan fingerprints in the 2D CNN then that should be emphasized more clearly.

[1] Kortvelesy, Ryan, Steven Morad, and Amanda Prorok. "Generalised f-mean aggregation for graph neural networks." Advances in Neural Information Processing Systems 36 (2023): 34439-34450. [2] Wu, Ang-Kun, et al. "Compact Spin-Charge Separated Neural Quantum States for Valence-Bond States." arXiv preprint arXiv:2606.17045 (2026).

Questions for authors

Can you clarify the difference between the GMP operation used in [2] to the one you are introducing here? For the lipophilicity task, can you evaluate the models from figures 5 and 6 on the same scale so they will be directly comparable? In Figure 5, is the advantage of the local max with global geo statistically significant compared to the other local max models?

Minor: In figures 4 and 5 the error bars are cut off by the upper/lower limit of the y-axis. Some acronyms are undefined (RG) or used before they are defined (GMP) in the main text.

References: [1] Kortvelesy, Ryan, Steven Morad, and Amanda Prorok. "Generalised f-mean aggregation for graph neural networks." Advances in Neural Information Processing Systems 36 (2023): 34439-34450. [2] Wu, Ang-Kun, et al. "Compact Spin-Charge Separated Neural Quantum States for Valence-Bond States." arXiv preprint arXiv:2606.17045 (2026).

3 comments
AuthorsOpenReview ↗

Response to Reviewer Wd85

Thank you for highlighting the prior-work relationship, sign controls and molecular comparability. For GMP, we clarify prior work, add sign controls and correct molecular evaluation.

Q1 / novelty concern: How does GMP differ from the motivating quantum-state paper?

Compact Spin-Charge Separated Neural Quantum States for Valence-Bond States (Wu et al., 2026; your [2]) provides the original physical motivation for GMP. It studies quantum-state representation using the same signed global operation, as stated in the Introduction.

Here we study the machine-learning aspects of GMP as a pooling/coarse-graining primitive: local/global placement, learned features and training behavior across controlled, image, molecular and physical tasks. Composition analysis identifies preserved statistics; the quantum benchmark uses Hamiltonian-defined targets. Sections 2–3 discuss generalized mean pooling (GeM), generalized f-mean aggregation (GenAgg) and renormalization-group (RG) analysis. GM denotes unsigned geometric mean pooling.

Weakness: Missing unsigned comparisons and the role of the sign term.

At n = 1000, signed/unsigned balanced accuracy is 100.00%/49.77% on A and 94.32%/50.63% on B; positive R1 gives identical mean squared error (MSE). Table 1; Appendix F, Table 9.

We also learn signed coefficients of up/down configurations in the lowest-energy state of an open 16-spin chain. Numerical diagonalization supplies labels; J2 couples sites two steps apart, with adjacent-site coupling 1.

ModelOverlap, J2 = 0.7Overlap, J2 = 1.0
Signed GMP0.9878 ± 0.00400.9258 ± 0.0147
Unsigned geometric mean (GM)−0.0336 ± 0.0402−0.0616 ± 0.0920
Learned log/phase0.9952 ± 0.00260.9761 ± 0.0070

Overlap is held-out cosine similarity, mean ± descriptive standard deviation (SD) over three splits and three initializations per split. Models share a convolutional neural network (CNN) and comparable head budgets; spin-flipped/reversed configurations stay together, selection uses validation, and sign calibration uses training. The signed/unsigned contrast supports useful sign preservation on physical data. Learned log/phase, an exponential/cosine wavefunction readout, reaches higher overlap. Section 4.5, Table 3; Appendix G, Table 10.

Weakness: With positive Softplus inputs, molecular GMP reduces to unsigned GM.

Indeed, positive softplus inputs make signed GMP equal unsigned GM. The molecular task therefore studies magnitudes and placement; the controlled and quantum ablations test the sign term.

Q2 / molecular-baseline concern: Same-scale evaluation and purpose of the baselines?

Replacement Table 2 compares all seven models on original logD units using shared Morgan fingerprints, five scaffold partitions (molecules sharing a framework stay together) and three runs per partition. Direct-logD root mean squared error (RMSE) is 0.866 for max/max and 0.887 for max/GM; max/GM versus max/average is unresolved (paired difference +0.00108 [−0.00444, 0.00660]). CNN labels denote local/global pooling.

Fingerprint baselines give reference performance without CNN pooling on the same inputs/splits; CNN variants isolate local/global aggregation.

We also test the base-10 inverse, 10^y. Max/GM has lower recovered-logD RMSE than max/max (1.126 versus 1.214), although both remain worse than direct-logD training. Section 4.4; Appendix E, Tables 7–8 gives paired comparisons and separates transformed-target scores from scores mapped back to logD.

Q3 / statistical-support concern: Is the apparent local-max/global-GM advantage supported?

The expanded, validation-selected evaluation does not support the apparent exp(y) advantage of global GM. For max/GM minus max/max, the transformed-target test R² difference is −0.02738, with descriptive 95% t interval [−0.04800, −0.00676]. The single-partition advantage is not reproduced under this common protocol.

Runs are averaged within each partition; intervals use the five partition means and describe sensitivity to these overlapping partitions.

Weakness: Image-classification performance.

Image/molecular studies explore transfer from the motivating scientific multiplicative setting. The CIFAR activation/clamp study raises ReLU global GMP from 82.49% to 84.54% with epsilon 10⁻³, versus 84.34% for matched average pooling (Appendix D.1, Table 6). This characterizes training sensitivity without claiming broad superiority.

Minor comments: Clipped error bars and undefined acronyms.

Plots show full SD whiskers; captions define seeds, symbols, scales and selection rules. We checked acronyms throughout the manuscript and captions and expanded them at first use, including GM, SD, GMP and RG.

Do the clarified contribution, sign controls and corrected molecular analysis address your main concerns?

Thank you to the authors for their detailed response and for running the additional experiments. The relationship to Wu et al., 2026 is clear, the new spin chain prediction supports the need of the sign term on a non synthetic task, and the molecular regression task is now standardized to the original log scale.

The overall contribution remains marginal as the empirical results are insufficient, especially given the pooling operation itself is not new (Wu et al., 2026). These are my major remaining concerns:

  1. The experimental evidence on non synthetic data is still limited. In the molecular regression task, signed GMP is not evaluated and the unsigned variant does not show a clear advantage compared with max-avg pooling baselines. The spin chain task does show the need for the signed term, but it is just a single example and the log/phase model outperforms it. The baseline pooling operations (max / avg) are not evaluated for this task. This is not enough to establish the broader benefit of the method. In the authors’ view, based on the results in this work, what is the use case for GMP?
  2. The empirical characterization remains underdeveloped. Why was image classification chosen to show the sensitivity to activation, clamp, and placement choices? Can the authors show or clarify which conclusions they expect to generalize to tasks where GMP would be applied? Using Softplus activation in the molecular regression experiment makes it unclear whether the conclusions generalize to the signed operation. To a lesser extent this also applies to the image classification experiments with the ReLU activation, which removes the negative sign. The characterization results would be stronger with activations that retain negative values (like tanh used in the spin chain task) and on tasks where GMP would realistically be applied.
  3. In Figure 4, the caption reads “the reported score is the maximum accuracy on the official test set across training epochs”. Can the authors clarify the way the tested checkpoint was chosen? Was it not not chosen based on the validation set?

Other minor concerns:

  • It should be stated clearly in the manuscript that GMP operation has been used in previous work and the objective of this work is to study the machine learning aspects.
  • In the QUANTUM MANY-BODY AMPLITUDE LEARNING section, SzS^z and ψ^\hat{\psi} are not explicitly defined in the main text.
AuthorsOpenReview ↗

Response to the remaining concerns and updates in the revised manuscript

Thank you for the follow-up. We have added matched pooling baselines, signed-feature controls and an exact singlet-product benchmark. The molecular numbers below supersede those in our earlier response: standardized-target fits have been replaced by raw-target experiments.

1. Intended use case and contribution.

GMP supplies a parameter-free prior when local factors combine multiplicatively and their sign parity matters. Singlet-product amplitudes provide a concrete physical example: local rules take values {−1, 0, +1}, and their product gives the amplitude up to normalization. We now test this setting at the exact J2 = 0.5 point of the uniform J1–J2 chain (J1 = 1). The manuscript explicitly credits Wu et al. (2026) for using the signed operation; our contribution studies its machine-learning behavior. We do not claim broad superiority on image or molecular tasks. Our results support GMP as a structural prior for problems with known signed multiplicative structure, illustrated here by quantum-state amplitudes and controlled synthetic benchmarks.

2. Quantum pooling baselines and the log/phase alternative.

All methods use the same tanh backbone, comparable head budgets, 2000 updates per learning rate and validation-only checkpoint/rate selection. Held-out signed overlaps are:

ModelJ2 = 0.5J2 = 0.7J2 = 1
Signed GMP0.999875 ± 0.0001080.9878 ± 0.00400.9258 ± 0.0147
Unsigned GM−0.0472 ± 0.0613−0.0336 ± 0.0402−0.0616 ± 0.0920
Average0.7119 ± 0.29440.9237 ± 0.08170.7944 ± 0.1869
Max−0.0637 ± 0.20060.3288 ± 0.24580.2775 ± 0.1798
Log/phase0.9571 ± 0.12860.9952 ± 0.00260.9761 ± 0.0070

Entries are mean ± sample SD over three configuration splits and three initializations per split (Section 4.5, Table 2; Appendix G, Tables 13–15).

At J2 = 0.5, GMP reaches overlap 0.999706–0.999976 and 100% Born-weighted sign accuracy in every run, exceeding average in all nine matched runs. This supports stable learning of a physically defined singlet-product state. The three-split GMP-minus-average interval remains broad, +0.2880 [−0.2726, 0.8486], and the corresponding intervals at J2 = 0.7 and 1 also include zero; we do not claim statistical significance. The signed/unsigned contrast supports preserving signs in this architecture.

Log/phase uses a learned cosine phase modulation multiplying a positive envelope. It reaches near-unit overlap in eight of nine J2 = 0.5 runs; one reaches 0.6141, accounting for its lower mean. It achieves higher mean overlap at J2 = 0.7 and 1. We retain it as a useful alternative, not a resolved GMP win. The evidence covers one spin-chain family.

3. Molecular sign controls and the size of the benefit.

You are correct that Softplus removes negative features, so that experiment tests geometric magnitudes rather than sign preservation. Appendix E also compares signed GMP, unsigned GM, average and max with tanh, fixing local max. Targets are raw y = log10(D), auxiliary exp(y), or physical D = 10^y, without shifting or scaling.

Figure 5 evaluates nine placements at epoch 200 over ten seeds on one scaffold partition: local-max/global-GMP gives R² = 0.3545 versus 0.3466/0.3460 for global max/average, a marginal lead. The matched comparisons use five scaffold partitions, three seeds and minimum native-target validation RMSE. On raw exp(y), Softplus GMP/average R² values are essentially tied (0.3049/0.3056), versus 0.2697 for max. With tanh, signed GMP is below average/max on exp(y) (0.2668 vs 0.2958/0.2987) and raw y (0.4294 vs 0.4920/0.4985), though above unsigned GM on raw y (0.3506). These controls do not establish a general molecular advantage (Tables 6–11). Both partitioning and selection differ from the placement study.

4. Characterization and generalization.

Images serve as familiar learned-feature stress tests of placement and activation compatibility. Sign removal, exact zeros and small magnitudes affect GMP across tasks; optimal activations, clamps and placements remain task-dependent. We added tanh molecular controls and tanh/LeakyReLU activation/clamp controls on the physical task. Increasing the tanh GMP clamp from 10^−6 to 10^−3 changes overlap at J2 = 1 from 0.9258 to 0.9588. Unsigned GM partially recovers with LeakyReLU, showing representation dependence (Appendix G, Table 14).

5. Figure 4 selection and notation.

Figure 4 reports each run's maximum official-test accuracy across epochs, without independent validation selection; we now explicitly identify these as optimistic test-set maxima. Appendix D.1, Table 5 instead uses a held-out training split for checkpoint selection. Section 4.5 defines total magnetization S^z_tot and the prediction ψ̂(s). The molecular figures display full SD whiskers.

Thank you again for your careful feedback. We hope the added controls and clarified scope make the contribution and its limitations clearer.

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

The paper proposes Geometric Mean Pooling (GMP). GMP multiplies the signs of the inputs in a window and scales that by the geometric mean of their clamped magnitudes. The authors motivate it with renormalization-group (RG) coarse-graining and quantum many-body states. They show that non-overlapping hierarchical GMP gives the same result as a single global GMP. The evaluation covers:

  • synthetic sign-product classification (Cases A and B)
  • synthetic geometric-mean regression (R1 to R4)
  • MNIST, Fashion-MNIST and CIFAR-10
  • MoleculeNet Lipophilicity

The authors frame GMP as a regime-dependent inductive bias, not a general replacement for average or max pooling.

Critical strengths

The paper is honestly scoped. It presents GMP as a regime-dependent inductive bias, not a universal replacement, reports negative results openly, and changes only the pooling primitive within each experiment. The sign-parity ablation cleanly shows what the sign term contributes.

The operator is simple, parameter-free, and easy to implement, and the code is released. The hierarchical consistency property is stated precisely, together with the conditions it requires.

Critical weaknesses

The synthetic evidence is circular. The Case A and B labels and the R1/R2 targets are defined by the GMP formula itself. The average and max baselines are followed by only a linear layer, so they cannot represent parity or log-magnitude statistics. R4 scores every method with a geometric-mean readout that GMP preserves by construction. Baselines such as log-features with average pooling, GeM, or NPU are missing, so the results show the construction works rather than an advantage in learning.

The real-data results do not support the method. GMP clearly underperforms on CIFAR-10, and a positive activation that could avoid the ReLU-zero problem was never tested. In the lipophilicity experiment, the target transform uses exp instead of 10^ for a base-10 log D. Global GMP wins only in one setting, and only within the error bars. On the original log scale, max pooling performs best.

Questions for authors
  1. How does GMP compare with log|x| and negative-sign features followed by average pooling and an MLP, and with GeM and NPU, on Cases A/B and R1–R3?
  2. Is z = exp(y) intentional, given that log D is base-10? What happens with z = 10^y, and with the CNN evaluated on the log scale next to the Table 2 baselines?
  3. Do the image results improve with softplus or leaky-ReLU backbones, or with an unsigned GMP and a larger ε?
  4. Why do average and max pooling give the same 60.4% at ρ = 0.9 in Case A?
  5. Can you show a non-synthetic task where GMP clearly beats the additive-in-log-space alternative?
2 comments
AuthorsOpenReview ↗

Response to Reviewer NRYj

Thank you for identifying the missing controls. We added the requested controls and a physical benchmark for geometric mean pooling (GMP).

Q1 / Weakness 1: Stronger log/sign, generalized-mean and arithmetic baselines on A/B and R1–R3.

Analytic log-magnitude and parity reconstruct GMP exactly (with the same clamp/zero rule). We study the learning benefit of encoding a known multiplicative rule, rather than representational impossibility of alternatives.

Methods share a training budget and validation selection across five seeds. We compare unsigned geometric mean (GM), magnitude-based generalized mean pooling (GeM) and gated real Neural Power Unit (RealNPU) summaries with multilayer perceptron (MLP) readouts. At n = 1000:

MethodCase A (%)Case B (%)
Average + MLP50.0452.05
Max + MLP49.6955.40
Signed GMP + MLP100.0094.32
Unsigned GM + MLP49.7750.63
Log/sign mean + MLP53.5050.24
Log/sign count + MLP60.5651.39
Deep Sets mean + count50.3691.15
Magnitude GeM + MLP50.0349.82
Gated RealNPU + MLP49.9750.55

Entries are mean test balanced accuracy (average class recall). A predicts global parity; B predicts the sign of summed clean cellwise signed geometric summaries from noisy inputs.

Deep Sets approaches GMP on B as data increase, supporting sample efficiency over an expressivity claim. Regression rankings vary: multiplicative-noise R3 mean squared error (MSE) is 6.57/6.58 × 10⁻³ for GMP/log-sign MLP; additive-noise Deep Sets improves on GMP (4.34 versus 6.49 × 10⁻⁴). Appendix F, Table 9 gives R1–R3; Figure 7 shows sample-size and mixed-target results. Here pooling acts directly on raw inputs. R4 tests composition with the common geometric readout stated in Figure 3.

Q2 / Weakness 2: Why exp(y), and what happens with 10^y and common-scale baselines?

exp(y) is an auxiliary reparameterization, not the inverse of base-10 logD. Both use matching inverses and recovered-logD validation root mean squared error (RMSE) for selection. Test-logD RMSE for the convolutional neural network (CNN) variants is:

Training targetCNN max/maxCNN max/GM
Original logD0.8660.887
exp(y)0.9380.953
10^y1.2141.126

Names denote local/global pooling. Means use five shared scaffold partitions, three runs each. Base-10 training changes the pooling ranking but does not improve upon direct-logD training. Table 2 compares all CNN/fingerprint baselines in original logD units; Appendix E, Tables 7–8 gives paired intervals, transformed-target scores and the inverse policy.

Q3 / Weakness 2: Do activation, unsigned pooling and larger epsilon help?

Larger epsilon substantially improves GMP. Across 18 global-CIFAR conditions and three seeds, raising epsilon from 10⁻¹² to 10⁻³ improves GMP from 82.49% to 84.54% with ReLU and 73.21% to 84.87% with leaky-ReLU; matched average pooling gives 84.34%/85.25%. Softplus GMP gives 80.05%, versus 79.58% for average. Unsigned GM at 10⁻³ gives 83.58%/84.45%/79.92% for ReLU/leaky-ReLU/softplus. Models share architecture, augmentation and budget; checkpoints use training-derived validation data. Exact zeros still set signed GMP to zero. Appendix D.1, Table 6 reports all conditions and across-seed standard deviation (SD); Figure 4 separately states its epoch-maximum test metric.

Q4: Why do average and max both give 60.4% at rho = 0.9?

Both predict class zero for every example; 60.4% is that test set's majority-class fraction, so balanced accuracy is 50%. Section 4.1; Appendix C.1. Numerical underflow affected 17/1000 test labels in the length-128 rerun. That row now uses stable integer-parity test labels; all new experiments use integer parity (Appendix H).

Q5: Is there a physical task where GMP beats an additive-in-log-space alternative?

The answer depends on the alternative. On the open 16-spin J1–J2 chain, GMP strongly improves over log/sign counts but does not outperform learned log/phase:

ModelOverlap, J2 = 0.7Overlap, J2 = 1.0
Signed GMP0.98780.9258
Log/sign counts0.1567−0.0548
Learned log/phase0.99520.9761

Targets are signed coefficients of up/down configurations in the lowest-energy state, obtained by numerical diagonalization. J2 couples sites two steps apart; adjacent coupling is 1. Overlap is held-out cosine similarity, averaged over three splits and three initializations. Models share a CNN and comparable head budgets, with validation selection and train-only sign alignment. This supports useful fixed signed pooling; learned log/phase is stronger. Section 4.5, Table 3; Appendix G, Table 10 reports SDs and all models.

Do these controls and the physical benchmark address your main concerns about the synthetic comparisons and evaluation?

AuthorsOpenReview ↗

Correction to Q2 molecular evaluation and update to the physical benchmark

Thank you again for identifying the evaluation issues. We have uploaded a revised manuscript and need to correct Q2 of our earlier response.

The earlier Q2 molecular table is superseded. Its numbers used standardized-target fits and selection by recovered-logD validation RMSE. The replacement experiments instead train directly on raw y = log10(D), raw exp(y), or raw D = 10^y, without shifting, centering, standardization or rescaling. Every model, including Morgan FC, RF and XGBoost, is refitted for each target. Neural models minimize native-target MSE for 200 epochs and retain the checkpoint with minimum validation RMSE in that same native-target space. Evaluation on recovered logD is secondary and does not select the checkpoint.

Appendix E, Tables 6–8 give the replacement native-space results. For example, with local max fixed, test RMSE/R² means are:

Raw targetFeature activationSigned GMPAverageMax
exp(y)Tanh13.0963 / 0.266812.8349 / 0.295812.8095 / 0.2987
exp(y)Softplus12.7524 / 0.304912.7472 / 0.305613.0668 / 0.2697
y = log10(D)Tanh0.9194 / 0.42940.8670 / 0.49200.8614 / 0.4985
y = log10(D)Softplus0.8885 / 0.46730.8938 / 0.46140.8867 / 0.4691

Means average three runs within each of five shared scaffold partitions, then average partition means. Native RMSE and R² are interpreted within each target space, not across nonlinear transformations. On raw y, Morgan FC, RF and XGBoost give R² = 0.4357, 0.4411 and 0.4247, respectively (Table 7). Raw-10^y results are retained in Table 8; they do not establish a general GMP advantage.

exp(y) is intentional as an auxiliary target, but it is not the physical inverse of log10(D). For secondary inverse evaluation we use ln(ẑ) or log10(D̂), respectively. No prediction flooring, shifting, absolute value or removal of nonpositive observations is used. If any prediction in a test fold is nonpositive, that run's inverse metric is undefined; a 15-run aggregate is given only when every run is valid (Table 9). Thus the complete recovered-logD RMSE table in our earlier Q2 should not be used for the new protocol.

Section 4.4 retains the original Figure 5: the raw-exp(y), nine-placement experiment, evaluated at epoch 200 over ten seeds on one scaffold partition. The GMP lead there is marginal. Later comparisons, paired intervals and final-epoch sensitivity are separated in Appendix E, Tables 6–11. Both the partitions and checkpoint rule differ from that original sweep.

Physical benchmark update (Q5). We evaluate the exact singlet-product point J2 = 0.5 of the uniform J1–J2 chain (J1 = 1), where amplitudes multiply local {−1, 0, +1} singlet rules up to normalization. This supplies a physical test of the multiplicative prior. Table 2 retains J2 = 0.7 and 1 and includes matched average/max baselines at all three couplings.

At J2 = 0.5, mean ± SD held-out overlaps are 0.999875 ± 0.000108 (signed GMP), −0.0472 ± 0.0613 (unsigned GM), 0.7119 ± 0.2944 (average), and −0.0637 ± 0.2006 (max). All use the same tanh backbone, comparable head budgets, three splits, three initializations and a three-rate validation search with 2000 updates per candidate. GMP reaches near-unit overlap and 100% Born-weighted sign accuracy in all nine runs, exceeding average in every matched run. The split-level GMP-minus-average interval remains +0.2880 [−0.2726, 0.8486]. At J2 = 0.7 and 1, average reaches 0.9237 and 0.7944 (GMP 0.9878 and 0.9258); the paired GMP-minus-average intervals include zero at all three couplings.

Log/phase reaches near-unit overlap in eight of nine runs at this point; one reaches 0.6141, giving mean 0.9571 ± 0.1286. The GMP-minus-log/phase interval is +0.0428 [−0.1417, 0.2272]. Thus the new result supports stable learning with GMP, without resolving superiority over the learned alternative. Log/phase still has higher mean overlap at J2 = 0.7 and 1. Appendix G, Tables 13–15 give all models, activation/clamp controls and the exact-point results.

The learned synthetic comparisons (now Table 12), mixed-target experiment and validation-selected CIFAR controls from our earlier response remain unchanged. We have also clarified that original Figure 4 uses epoch-maximum official-test accuracy, whereas Appendix D.1, Table 5 uses validation-selected checkpoints.

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

The paper proposes signed geometric mean pooling (GMP) within each pooling window. The output of GMP is the product of input signs times the geometric mean of the magnitudes. Because GMP is an arithmetic mean in log-magnitude space, nested equal-size non-overlapping GMP equals global GMP, which the authors present as a renormalization-group fixed-point property. Experiments cover synthetic sequence tasks, image classification on MNIST/Fashion-MNIST/CIFAR-10, and molecular lipophilicity regression. GMP is best on the synthetic tasks, which are designed to match the operator. On images and molecules it is mostly worse than or comparable to standard pooling.

Critical strengths
  1. The operator is simple, parameter-free, cheap to compute in log space, and clearly defined (Eq. 1).
  2. The paper openly states that it does not propose a general replacement for standard pooling, and discusses its limitations.
Critical weaknesses
  1. For positive inputs, the operator is the p→0 limit of GeM pooling, and its magnitude part is GenAgg with f = log|·|, as the related-work section notes. The novelty is the addition of sign information, and it brings no benefit on real data.
  2. The only clear benefit in the numerical results comes from synthetic datasets designed to match the operator.
  3. The real-data results do not support GMP. It is clearly worse on CIFAR-10. On lipophilicity it wins only after the target is changed from logD to exp(logD); the baselines in Table 2 are reported on the original log scale, so the numbers cannot be compared, and on the original scale max pooling is best.
Questions for authors
  1. What happens when the target is a mix of additive and multiplicative terms?
  2. Is there a task from the quantum many-body literature that you could use to validate your approach?
2 comments
AuthorsOpenReview ↗

Response to Reviewer pnCY

Thank you for the constructive suggestions. We added both requested experiments: mixed additive/multiplicative targets and a quantum many-body task with labels obtained independently by numerical diagonalization. We also added stronger learned baselines and corrected the molecular comparison.

Questions

Q1: “What happens when the target is a mix of additive and multiplicative terms?”

Neither summary is uniformly best: when both structures contribute, combining them recovers the target. We interpolate between arithmetic-mean and signed geometric mean pooling (GMP) components on Gaussian sequences, standardizing each using reference means and standard deviations (SD) kept fixed across splits. Average pooling and GMP trade performance as the multiplicative weight changes. A hybrid receiving both summaries achieves mean test R² > 0.9999 at weights 0.25, 0.5 and 0.75 with 1000 training examples. As expected from the target construction, this illustrates how additive and multiplicative aggregation complement each other. Appendix F.2; Figure 7, right.

Q2: “Is there a task from the quantum many-body literature?”

Signed GMP learns physical amplitudes and their sign structure. We predict the signed coefficients of up/down spin configurations in the lowest-energy state of an open 16-spin J1–J2 chain. Adjacent-site coupling is fixed to 1; J2 controls interactions between sites two steps apart. Numerical diagonalization supplies the targets independently of GMP.

ModelOverlap, J2 = 0.7Overlap, J2 = 1.0
Signed GMP0.9878 ± 0.00400.9258 ± 0.0147
Unsigned geometric mean (GM)−0.0336 ± 0.0402−0.0616 ± 0.0920
Learned log/phase0.9952 ± 0.00260.9761 ± 0.0070

Overlap is cosine similarity of held-out coefficient vectors. Entries are mean ± descriptive SD over three splits and three initializations per split. Models share a convolutional backbone and comparable head budgets, with validation selection and one overall sign fitted on training data.

The signed/unsigned contrast demonstrates that retaining sign information is important in this physical task. The learned log/phase readout reaches higher overlap; it forms the wavefunction through a learned exponential/cosine rule, whereas GMP supplies fixed pooling before a multilayer perceptron. Section 4.5, Table 3; Appendix G, Table 10.

Critical weaknesses

Weakness 1: Established positive-input pooling and no sign benefit on real data.

Positive geometric means connect to generalized mean pooling (GeM) and generalized f-mean aggregation (GenAgg). The cited valence-bond-state paper uses the same signed global operation in a quantum-state representation and provides the physical motivation for this study. The GMP paper focuses on the machine-learning aspects of signed multiplicative pooling: local/global placement, learned representations, training behavior, and evaluation across controlled, image, molecular, and physical settings. The composition analysis and controlled tasks characterize this inductive bias. The physical signed/unsigned comparison in Q2 directly tests the sign contribution.

Weakness 2: Synthetic tasks are designed to match GMP.

These tasks isolate known composition rules. The stronger baselines test learning efficiency, rather than unique representability. At n = 1000, Case B balanced accuracy is 94.32% for GMP and 91.15% for Deep Sets mean-plus-count. The gap to the validation-selected learned comparison shrinks from 6.83 to 1.34 percentage points as n grows from 250 to 4000. This supports an inductive-bias/sample-efficiency interpretation. Appendix F, Table 9; Figure 7, left.

Weakness 3: CIFAR underperformance and incomparable molecular scores.

The reported CIFAR placement results favor standard pooling; the activation/clamp controls are in Appendix D.1, Table 6.

Table 2 now compares all seven models on original logD units under shared scaffold partitions (molecules with the same ring-and-linker framework stay together). Max/max remains better than max/GM on direct-logD training, and the expanded analysis does not support the apparent exp(y) advantage. Appendix E separates transformed-target and recovered-logD scores. Generic image/molecular studies explore applicability; clear scientific multiplicative structure remains GMP's motivating setting.

Do the mixed-target and quantum experiments address the two questions raised in your review? We would welcome your assessment.

AuthorsOpenReview ↗

Correction to our molecular results and update on the quantum comparison

Title: Correction to our molecular results and update on the quantum comparison

Thank you again for your suggestions. We have uploaded a revised manuscript and would like to correct the molecular portion of our earlier response and update the quantum comparison.

Correction to the molecular results. The molecular comparison in the “Critical weaknesses, Weakness 3” paragraph of our earlier reply (the seven-model Table 2 and the max/max versus max/GM statement) came from a protocol using standardized targets. We have replaced those fits with experiments on raw targets and changed checkpoint selection to match the training objective. That earlier comparison is superseded by Appendix E, Tables 6–11.

In the replacement comparisons, every model—including Morgan FC, RF and XGBoost—is fitted directly to raw y = log10(D), exp(y), or 10^y, without target shifting or scaling. Neural models minimize MSE and select the checkpoint with lowest validation RMSE in that same target space. Five scaffold partitions and three seeds per partition are shared across models. CNNs fix local max and vary global pooling, with both tanh and Softplus features.

For raw exp(y) with Softplus, GMP and average give nearly identical native-space R² (0.3049 and 0.3056), versus 0.2697 for max. For direct raw-logD training with tanh, GMP gives RMSE/R² = 0.9194/0.4294, versus 0.8670/0.4920 for average and 0.8614/0.4985 for max. These results do not establish a general molecular advantage for GMP.

Section 4.4 retains the original nine-placement Figure 5, evaluated at epoch 200 over ten initialization seeds on one scaffold partition. Its raw-exp(y) local-max/global-GMP mean R² = 0.3545 is only marginally higher than global max/average (0.3466/0.3460). The later matched comparisons are reported separately in Appendix E; they change both partitioning and selection. Scores from different nonlinear target spaces are not treated as directly comparable. exp(y) is explicitly an auxiliary target; the physical inverse is D = 10^y.

Quantum update. Section 4.5, Table 2 now includes the exact singlet-product point J2 = 0.5 of the uniform J1–J2 chain (J1 = 1), alongside J2 = 0.7 and 1. At the exact point, amplitudes multiply local {−1, 0, +1} singlet rules, up to normalization, providing a physical test of the signed multiplicative prior.

With matched tanh backbones, comparable head budgets and validation searches, mean ± SD held-out overlaps at J2 = 0.5 are 0.999875 ± 0.000108 for signed GMP, −0.0472 ± 0.0613 for unsigned GM, 0.7119 ± 0.2944 for average, and −0.0637 ± 0.2006 for max. Each mean uses three splits and three initializations per split, with 2000 updates at each of three learning rates. GMP reaches near-unit overlap and 100% Born-weighted sign accuracy in every run, exceeding average in all nine matched runs. The three-split GMP-minus-average interval is +0.2880 [−0.2726, 0.8486], so we do not claim statistical significance. At J2 = 0.7 and 1, average reaches 0.9237 and 0.7944 (GMP 0.9878 and 0.9258); the paired GMP-minus-average intervals include zero at all three couplings.

Learned log/phase reaches near-unit overlap in eight of nine J2 = 0.5 runs; one reaches 0.6141, giving mean 0.9571 ± 0.1286. It remains stronger in mean overlap at J2 = 0.7 and 1 (0.9952/0.9761 versus GMP 0.9878/0.9258). Its definition and results remain in Appendix G: Table 13 covers J2 = 0.7/1, Table 14 activation/clamp controls, and Table 15 the exact point. This supports GMP's stable learning at the physically matched point without establishing general superiority over learned phase composition.

The mixed-target and learned-baseline results discussed in our earlier response are unchanged (Appendix F; the learned-baseline table is now Table 12). The CIFAR control table previously cited as Table 6 is now Table 5. Our conclusion remains that GMP supplies a structural prior for suitable signed/multiplicative settings, rather than a generally superior pooling operation.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.