acceptodds
Under review as a conference paper at ICLR 2027

Feature Correlation Preserving Quantization for Analog Compute-in-Memory

Abstract

Analog compute-in-memory (ACIM) enables efficient neural network inference by performing matrix operations within memory arrays, but its limited numerical precision requires low-bit quantization. Existing post-training quantization (PTQ) methods mainly minimize reconstruction error between full-precision and quantized representations. This objective does not explicitly preserve feature relations, which can be important when multiple features share the same quantization configuration within an ACIM mapping group. We propose Feature Correlation Preserving Quantization (FCPQ), a groupwise PTQ framework that preserves feature relations while optimizing reconstruction. FCPQ consists of two components. First, Feature Correlation Calibration (FCC) standardizes features and aligns the correlation matrices of full-precision and quantized representations within each mapping group. Second, Correlation Aware Quantization Optimization (CQO) estimates the contribution and distortion of feature relations and uses them to assign adaptive reconstruction weights. Together, these components guide quantization toward preserving both numerical values and feature relations. Experiments across multiple datasets and quantization settings show that FCPQ improves model accuracy and maintains the feature structure of full-precision models under low-bit quantization.

Primary area:
optimization
Keywords:
Low-Bit Quantization, Feature Correlation, Analog Compute-in-Memory, Efficient Inference

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_w3FEOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper presents a method called FCPQ for low precision post training quantization for SRAM-based ACIM. The method is built on a hardware constraint where the quantization parameters need to be shared across output features. This is a constraint that does not exist in GPU settings and is not addressed in most PTQ methods. The main contribution of the paper is the FCPQ method which has two key components:

  1. Feature Correlation Calibration (FCC): The method calculates normalized vectors per output channel. Then it calculates the dot product of each vector that is within the same group and share the quantization parameters. Then it creates this loss that tries to preserve the relations of the FP model to the quantized one.
  2. Correlation aware quantization optimization: Create a feature importance weight that scales the reconstruction objective giving more emphasis on features that are highly involved in the original relation structure and suffer larger relation changes after quantization. The paper does an evaluation on three different CNN architecture, three different quantization settings and on two benchmarks.
Critical strengths
  1. The setting is unexplored: PTQ method under hardware-imposed scale sharing across output channels in a problem not addressed by PTQ methods.
  2. The method is cheap to apply and works with just a calibration dataset. The direction of the results is consistent and the method seems to outperform the other baseline methods.
Critical weaknesses

Critical Weaknesses:

  1. Section 4.5 is misleading. The text describes the macro as “fabrication-ready, post-layout,” while the caption of Table 5 reports “measured” accuracy on a “fabricated” chip. As a result these experiments are done in simulation something that is not mentioned anywhere. The simulator and noise-injection model are not described. The first row of Table 5 is the same as the first row of Table 1, so the authors imply that there in zero effect of the SRAM-based ACIM hardware on accuracy when operates on nominal conditions. Furthermore it is unclear how much is the reduced resolution in “Reduced ADC resolution”, what is the Accum. noise, σ = 2% FS, what does it mean 10% IR-drop and what is the 5% drift on SRAM-based ACIM?

  2. The gains are not statistically supported and baseline comparison is unclear and incomplete. The gains that are presented in the tables of the paper are relatively small, as a results it is unclear if the method is consistently better or the improvement is within statistical error. The authors present results from two resent methods BOF4 and AnyBCQ that present results for LLMs. Especially the BOF4 method is not clear how it can be used for ACIM as it uses a codebook. There are methods that are missing from the comparison like PD-Quant and QDrop which have made for CNNs, reportedly beat BRECQ and the authors cite them in the related work.

Questions for authors
  1. Which framework did you use for the simulation, what is the noise model you used for evaluation and what are the details of the noise configurations on Table 5.
  2. Can you explain what does it mean 5% drift for an SRAM-based ACIM?
  3. What is the difference between the Adaptive rounding in section 3.2 and AdaRound? Adaround should be cited in section 3.2. In the AdaRound paper in equation (24) \beta is decreasing in this paper equation (7) \beta is increasing. Can you explain this difference?
  4. An equation with the total loss is missing in section 3.5. Do you use also the L_round that is specified in section 3.2?
  5. The method presented in FCC is a standard correlation metric, can you comment on the novelty of this contribution and if it has been used in other similar works?
  6. Can you elaborate if a codebook quantization method like BOF4 can be used in SRAM-based ACIM?
  7. Can you explain the process for obtaining the results for W3A3 and W2A4 results for BOF4, as the method is made for 4-bit quantization.
  8. The same hyperparameters have been used for the training of the all baseline methods. How are these hyperparameters have been selected? Are these the optimal hyperparameters for FCPQ. Can you demonstrate that these are the optimal hyperparameters for all the methods?
  9. On equation 18 you specify us, ue which define the activation interval. In the experiments this values are not specified. Can you explain how you specify these values and what values you use in the results.
4 comments

Dear Reviewer w3FE,

We sincerely thank you for recognizing the unexplored setting of PTQ under hardware-imposed scale sharing across output channels, as well as the practical value of a method that can be applied using only a calibration dataset. We also appreciate your positive assessment that the results are directionally consistent and that FCPQ outperforms the baseline methods. Thank you for your constructive comments and suggestions. Below, we provide detailed responses to address your concerns.

Q1: Clarification of Hardware Simulation and Noise Modeling

Response

We sincerely thank the reviewer for this valuable comment. To clarify the hardware evaluation, we separate post-layout circuit characterization, network-level error injection, and physical chip measurement.

We have revised Section 4.5 and the Table 5 caption to identify these results as a network-level evaluation based on post-layout macro characteristics, rather than as accuracy measured on a fabricated 22-nm chip.

(1) Evaluation scope and framework. To clarify which framework is used for the accuracy evaluation, we use PyTorch 2.4.1 with CUDA 12.4 to run ResNet-18/ImageNet-1K inference and inject the hardware effects before the corresponding conversion or accumulation step. The injected parameters come from the post-layout characterization of the 22-nm SRAM-based ACIM macro. The macro uses a nominal 0.8 V supply, 100 MHz clock, and a 4-bit SAR ADC. The nominal row is the unperturbed software reference with the nominal quantization setting.

To keep the comparison fixed, we use W4A4 ResNet-18 with mapping-group size 32, the same pretrained checkpoint, a 1,024-image calibration subset, an independent 4,096-image diagnostic set, and the complete ImageNet-1K validation set for final evaluation. Trainable quantization methods use 2,000 Adam iterations with batch size 64 and learning rate 10−210^{-2}; calibration indices, diagnostic indices, loader order, and evaluation split are shared across methods.

To correct potentially misleading experimental wording in the original paper, we have revised the first two rows of Table 5 as follows and made the same change in the paper text and caption.

Reference rowMin-Max Top-1 (%) ↑AnyBCQ Top-1 (%) ↑FCPQ Top-1 (%) ↑
Unperturbed software reference67.2469.3569.41
Nominal ACIM simulation66.8368.9369.01

The nominal simulation shows an accuracy drop for all three methods, clarifying the effect of ACIM hardware under nominal conditions.

To check whether FCPQ's advantage also holds in physical chip evaluation, we evaluate W4A4 ResNet-18 on ImageNet-1K using the ACIM platform used for chip-level evaluation, including an ACIM-based near-sensor neural network computing chip (NS-Engine). FCPQ achieves the highest chip-measured Top-1 among the five methods.

MethodChip-measured Top-1 (%) ↑
BRECQ59.87
AnyBCQ60.14
QDrop60.21
PD-Quant60.34
FCPQ62.78

(2) Noise and hardware configurations. To specify the conditions in Table 5, we evaluate each nonideality separately while keeping the model, quantization setting, data, and optimization protocol fixed. The low-supply condition uses 0.75 V. For reduced ADC resolution, we change the nominal 4-bit ADC to 3-bit and keep the same full-scale clipping. For accumulation noise, we add independent zero-mean Gaussian noise, n∼N(0,(0.02FS)2)n\sim\mathcal{N}(0,(0.02FS)^2), to the accumulated analog output immediately before ADC quantization. For IR drop, we apply a row-dependent linear attenuation that increases from 0 to 10% across 144 MAC rows. For effective conductance variation, we use the static cell-wise model specified below.

(3) Meaning of 5% variation. For the SRAM-based ACIM macro, 5% effective conductance variation represents a static cell-wise variation in the effective conductance relative to its programmed value G0G_0: G=G0(1+δ)G=G_0(1+\delta), where δ∼N(0,0.052)\delta\sim\mathcal{N}(0,0.05^2) is sampled once per effective cell, clipped to [−0.15,0.15][-0.15,0.15], and held fixed during inference.

Q2: Baseline Scope and Statistical Support

Response

We sincerely thank the reviewer for this valuable comment. To assess the consistency of the gains and complete the baseline comparison, we report matched CNN experiments and paired uncertainty estimates.

(1) Statistical support. To test whether FCPQ performs consistently across architectures and bit widths, we compare four architectures and 12 model-bit settings under matched checkpoints, calibration data, group size, optimization budget, and evaluation split. FCPQ ranks first in 11 settings; the only exception is a near-tie on ResNet-50 W4A4. For ResNet-18 W4A4, we use paired bootstrap on the same 50,000 ImageNet validation images with 10,000 resamples.

Paired comparisonDifference (pp)95% CI
FCPQ - AnyBCQ+0.06[-0.03, 0.15]
FCPQ - PD-Quant+0.11[0.02, 0.20]
FCPQ - QDrop+0.23[0.12, 0.34]

The intervals are above zero for PD-Quant and QDrop; the AnyBCQ comparison is close. The results across architectures and bit widths show that FCPQ's ranking is not limited to one setting.

(2) CNN baselines and matched tuning. To compare with the CNN-oriented methods cited in the paper, we evaluate QDrop and PD-Quant under the same checkpoint, W/A precision, mapping-group size, calibration data, evaluation split, and optimization budget. Each method retains its native objective and uses validation-selected parameters. QDrop selects its drop probability from {0.1, 0.2, 0.3, 0.4}; PD-Quant selects a multiplier from {0.25, 0.5, 1, 2} of its default prediction-difference coefficient.

On ImageNet-1K, FCPQ outperforms QDrop and PD-Quant in all 12 model and bit-width settings.

ModelBitQDropPD-QuantFCPQ
ResNet-18W4A469.1869.3269.41
ResNet-18W3A365.5366.1166.42
ResNet-18W2A456.5256.5358.49
ResNet-50W4A474.8275.0875.20
ResNet-50W3A371.2771.8472.80
ResNet-50W2A466.7666.9468.34
MobileNetV2W4A468.1568.2869.33
MobileNetV2W3A362.8663.1464.31
MobileNetV2W2A452.8254.5454.80
RepVGG-A0W4A468.9269.0769.44
RepVGG-A0W3A363.8364.1565.15
RepVGG-A0W2A454.5155.0255.81

To examine the comparison on another dataset, we use the same matched protocol on CIFAR-10. FCPQ achieves the highest Top-1 accuracy among the three methods in 11 of the 12 settings.

ModelBitQDropPD-QuantFCPQ
ResNet-18W4A493.7293.8794.13
ResNet-18W3A391.4791.6492.26
ResNet-18W2A484.9885.1786.41
ResNet-50W4A494.2694.3494.48
ResNet-50W3A392.3192.4593.04
ResNet-50W2A486.6386.8287.96
MobileNetV2W4A492.7492.8293.15
MobileNetV2W3A389.8890.0190.73
MobileNetV2W2A482.3282.5083.02
RepVGG-A0W4A492.9393.0692.97
RepVGG-A0W3A390.3590.4491.21
RepVGG-A0W2A482.3082.5183.85

(3) BOF4 as an algorithmic reference. We include BOF4 because it uses a low-bit PTQ parameterization different from conventional uniform quantization. Under the same CNN checkpoints, bit widths, and data, it serves as a useful algorithmic reference for assessing FCPQ on standard CNN quantization tasks and examining whether its advantage depends on a particular parameterization.

To support different weight precisions, we adapt the codebook size while retaining BOF4's block-wise normalization and codebook optimization procedure. Specifically, we independently optimize 16/8/4 codebook centroids for W4/W3/W2, respectively, and use the same specified A3/A4 activation quantizers as the other methods. W3 and W2 are thus lower-bit adaptations of the original BOF4 principle.

This comparison evaluates algorithmic accuracy only. We do not implement BOF4's codebook storage, decoding, or conductance mapping on ACIM. Its results should therefore not be interpreted as an ACIM hardware deployment or chip-level evaluation.

Q3: Relationship to AdaRound and Adaptive Rounding

Response

Thank you for asking us to clarify the relationship between FCPQ's adaptive rounding in Section 3.2 and AdaRound. FCPQ follows AdaRound's continuous relaxation, STE-based hard rounding, and rounding regularizer. The main difference is the β schedule: AdaRound decreases β, whereas FCPQ increases it linearly from 2 to 20 over 2,000 updates. To examine this difference, we conduct a controlled β-schedule ablation, keeping all other settings fixed.

β scheduleTop-1 (%) ↑Reconstruction error ↓
Fixed β=1069.170.0678
Increasing 2→2069.410.0649
Decreasing 20→269.220.0669
No rounding regularizer68.960.0721

The increasing schedule gives the best accuracy and lowest reconstruction error in this comparison, supporting its use under our calibration setting.

Q4: Complete Optimization Objective

Response

Thank you for asking us to clarify the complete optimization objective in Section 3.5. Yes. FCPQ includes the rounding regularizer defined in Section 3.2 at every calibration step. The complete scheduled loss is

Ltotal(u)=Lrec+ρ(u)[λcLFCC+α(LCQO−Lrec)]+Lround,u=t/T.L_{\mathrm{total}}(u)=L_{\mathrm{rec}}+\rho(u)\left[\lambda_cL_{\mathrm{FCC}}+\alpha\left(L_{\mathrm{CQO}}-L_{\mathrm{rec}}\right)\right]+L_{\mathrm{round}},\qquad u=t/T.

Here, ρ(u)\rho(u) gates only the relation terms; LroundL_{\mathrm{round}} remains always included during calibration. The component losses, LroundL_{\mathrm{round}}, and ρ(u)\rho(u) are defined in Section 3.

Q5:Novelty of Feature Correlation Constraint (FCC)

Response

(1) To our knowledge, this is the first groupwise PTQ method designed to explicitly preserve feature correlations under hardware-imposed shared-scale constraints in ACIM. Previous works have explored preserving relationships among feature representations in knowledge distillation [1; 2]. Meanwhile, existing PTQ methods, such as AdaRound and BRECQ, mainly focus on minimizing reconstruction errors [3; 4]. (2) ACIM quantization requires multiple output features within the same mapping group to share quantization parameters, coupling their quantization errors. This makes preserving relationships among jointly quantized features essential for maintaining the original feature structure. Furthermore, existing reconstruction-based PTQ approaches do not explicitly constrain these groupwise feature relationships. To address these issues, our FCC specifically aligns the correlation matrices of floating-point and quantized features within each ACIM mapping group, preserving feature relationships under hardware-imposed shared-scale constraints. (3) The results shown in the tables below demonstrate that FCC improves ImageNet Top-1 accuracy from 68.75% to 69.18% and reduces correlation distortion from 0.1028 to 0.0724. Furthermore, comparisons with alternative relation objectives show that FCC achieves the lowest Pearson distortion and the highest CKA score with comparable reconstruction errors, supporting its effectiveness in preserving feature relationships during hardware-aware quantization.

MethodFCCCQOTop-1 (%)Gain
Adaptive reconstruction✗✗68.75--
+ FCC✓✗69.18+0.43
+ CQO✗✓69.03+0.28
FCPQ✓✓69.41+0.66
Relation objectiveReconstruction MSE ↓Pearson distortion ↓ImageNet Top-1 (%) ↑CKA ↑
Covariance matching0.06880.083668.930.842
Variance-normalized MSE0.06840.065769.200.875
Pearson FCC0.06850.049869.410.916

[1] F. Tung and G. Mori, "Similarity-preserving knowledge distillation," in Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pp. 1365–1374, 2019. [2] J. Yim, D. Joo, J. Bae, and J. Kim, "A gift from knowledge distillation: Fast optimization, network minimization and transfer learning," in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp. 7130–7138, 2017. [3] M. Nagel, R. A. Amjad, M. van Baalen, C. Louizos, and T. Blankevoort, "Up or down? Adaptive rounding for post-training quantization," in Proceedings of the International Conference on Machine Learning (ICML), pp. 7197–7206, 2020. [4] Y. Li, R. Gong, X. Tan, Y. Yang, P. Hu, Q. Zhang, F. Yu, W. Wang, and S. Gu, "BRECQ: Pushing the limit of post-training quantization by block reconstruction," in Proceedings of the International Conference on Learning Representations (ICLR), 2021.

Q6: Hyperparameter Selection and Activation Interval Configuration

Response

Thank you for asking how the hyperparameters and activation interval are selected.

(1) Hyperparameter selection. For a fair comparison, we use the same checkpoints, W/A bit widths, mapping-group size, calibration data, and evaluation split. As stated in Section 4.1, all trainable methods use 2,000 Adam iterations, batch size 64, and learning rate 10−210^{-2} in the reported comparison. Each method retains its native objective and selects its method-specific parameters on the same held-out calibration split. The baseline search ranges and selected settings are:

MethodSearch rangeSelected setting
QDropDrop probability: {0.1, 0.2, 0.3, 0.4}0.2
PD-QuantPrediction-difference coefficient: {0.25, 0.5, 1, 2} × default1 × default
AdaRound / BRECQLearning rate: {10−310^{-3}, 10−210^{-2}}; native rounding or block parameters10−210^{-2}; validation-selected native settings

To examine whether conventional hyperparameter tuning explains FCPQ's gain, we compare default and tuned adaptive reconstruction with FCPQ under the same experimental conditions.

MethodOptimization strategyTop-1 (%)
Adaptive reconstructionDefault configuration68.75
Adaptive reconstructionTuned configuration68.81
FCPQFCC + CQO69.41

FCPQ remains ahead of the tuned reconstruction baseline, supporting the contribution of its relation-preserving design. The selected settings are the best within the tested ranges; they do not establish global optimality for FCPQ or every baseline.

(2) Activation interval. With normalized optimization progress u=t/Tu=t/T, usu_s and ueu_e define the relation-guidance window, not activation clipping values. Equation (18) sets ρ(u)\rho(u) to zero outside the interval and follows a half-sine within it. This lets reconstruction first establish numerical responses, then introduces relation guidance that returns to zero at the end of the window. We select FCPQ's parameters on the held-out calibration split, searching usu_s and ueu_e over the ranges below with ue>usu_e>u_s. The reported results use us=0.10u_s=0.10 and ue=0.90u_e=0.90, so relation guidance peaks at u=0.50u=0.50.

FCPQ parameterSearch rangeSelected value
λc\lambda_c{0.01, 0.03, 0.10, 0.30}0.10
α\alpha{0.25, 0.50, 0.75}0.50
τ\tau{0.05, 0.10, 0.20}0.10
usu_s{0.05, 0.10, 0.20}0.10
ueu_e{0.70, 0.80, 0.90}, with ue>usu_e>u_s0.90

To examine how the activation interval affects performance, we compare the selected window with alternative windows, including a full-duration window, under the same calibration setting.

Relation windowTop-1 (%) ↑Reconstruction error ↓Correlation distortion ↓
[0.00,1.00][0.00,1.00]69.310.06510.0537
[0.10,0.90][0.10,0.90]69.410.06490.0498
[0.20,0.80][0.20,0.80]69.330.06540.0521
[0.10,0.70][0.10,0.70]69.280.06570.0542

The selected window achieves the best accuracy and lowest errors in this comparison, supporting the chosen interval.

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

The work proposes Feature Correlation Preserving Quantization (FCPQ), a post-training quantization method that aims to not only recover features, as existing approaches do, but also the correlation of features. This is achieved by adding a loss function that penalizes different feature correlations between full-precision and quantized features and by adding a weighting scheme for reconstruction based on the relation participation and the relation distortion per feature. The approach is motivated by the usage of analog compute-in-memory (ACIM), which requires model quantization due to hardware limitations. The approach is experimentally benchmarked against several (classical and recently proposed) post-training quantization baselines on two datasets, showing that FCPQ outperforms the baselines in most of the presented scenarios. The experimental results are supplemented by an ablation study and an experiment on AIM compatible hardware.

Critical strengths
  • The proposed approach is overall well described, see the questions below for some exceptions.
  • The experimental setup is well documented. The results support the superior performance of FCPQ compared to several baselines. The ablation study confirms that all parts of FCPQ lead to an improved performance. The experiments include experiments using ACIM compatible hardware, showing that FCPQ also works in practice.
Critical weaknesses
  • The work points out that existing approaches do not explicitly preserve relations among features, but never give an underlying reason, why feature correlation preserving should increase model performance. A theoretical investigation would strengthen the motivation a lot.
Questions for authors

Clarification requests:

  • In Equation (3): Is sl.gs_{l.g} learnable or is rr the only learnable part of the quantization?
  • In Equation (17): Why is there a summand $\alpha(\mathcal{L}{CQO}-\mathcal{L}{rec})? Is the total loss Lfeat+Lrel\mathcal{L}_{feat} + \mathcal{L}_{rel}?
  • In Equation (18): Where does ρ(u)\rho(u) enters the training? Is ρ(u)\rho(u) a replacement for the constant α\alpha?
3 comments

Dear Reviewer UuKx,

We sincerely thank you for recognizing that the proposed approach is overall well described and that the experimental setup is well documented. We also appreciate your positive assessment that the results support FCPQ's superior performance compared with several baselines, that the ablation study confirms the contribution of all parts of FCPQ, and that the ACIM-compatible hardware experiments show that FCPQ works in practice. Thank you for your constructive comments and suggestions. Below, we provide detailed responses to your concern and questions.

Q1: Motivation for preserving feature correlations

Response

We sincerely thank the reviewer for this valuable comment. We agree that the motivation for preserving feature correlations should be supported by a clearer explanation of how correlation distortion affects downstream task performance. We therefore provide a theoretical analysis and controlled experiments to investigate this relationship.

(1) Theoretical motivation. To understand the effect of feature correlation distortion, we analyze how groupwise quantization errors alter the second-order statistics of feature responses. Let XfX_f and XqX_q denote the full-precision and quantized group responses, respectively, and write Xq=Xf+EX_q=X_f+E, where EE denotes the quantization error. The covariance shift caused by quantization is

ΔΣ=Cov⁡(Xf,E)+Cov⁡(E,Xf)+Cov⁡(E).\Delta\Sigma=\operatorname{Cov}(X_f,E)+\operatorname{Cov}(E,X_f)+\operatorname{Cov}(E).

A following linear map Y=XWY=XW transforms this change to

ΔΣY=W⊤ΔΣW.\Delta\Sigma_Y=W^\top\Delta\Sigma W.

Therefore, the orientation of relation changes relative to the downstream weights matters in addition to the total response error. Under a local second-order approximation of the downstream task loss, when the perturbation is small and its first-order expected contribution is negligible,

E[ΔL]≈12tr⁡ ⁣(HYW⊤ΣEW),\mathbb{E}[\Delta\mathcal{L}]\approx\frac{1}{2}\operatorname{tr}\!\left(H_YW^\top\Sigma_EW\right),

where HYH_Y is the Hessian of the task loss with respect to the downstream response and ΣE=Cov⁡(E)\Sigma_E=\operatorname{Cov}(E) is the quantization-error covariance. This expression shows that errors along task-sensitive feature directions can have a larger effect on downstream loss even when the total reconstruction error is similar. Standardized Pearson matching targets the pairwise structure, while LrecL_{\mathrm{rec}} preserves response means and magnitudes that are removed by standardization.

In the paper's W4A4 ResNet-18 setting, FCC measures standardized off-diagonal relations within each ACIM mapping group of 32 output features. The full FCPQ configuration uses λc=0.10\lambda_c=0.10, α=0.50\alpha=0.50, and CQO temperature τ=0.10\tau=0.10, with relation guidance scheduled over us=0.10u_s=0.10 to ue=0.90u_e=0.90.

(2) Controlled validation of feature relation preservation. To isolate the contribution of feature correlation preservation from numerical reconstruction, we conduct a controlled comparison using ResNet-18 on ImageNet-1K under W4A4 quantization with group size 32. We use the same pretrained checkpoint, calibration data, initialization, and optimization budget, disable CQO, match full-network reconstruction MSE within 0.5%, and change only the relation objective. As shown in the table, Pearson FCC lowers correlation distortion while improving ImageNet accuracy, CKA, and frozen-backbone transfer accuracy. The aligned gains indicate that relation preservation improves task-relevant representations when reconstruction quality is comparable.

A paired bootstrap on the same 50,000 ImageNet validation images with 10,000 resamples gives a +0.43-point Top-1 difference, with a 95% confidence interval of [0.31, 0.55].

Relation objectiveCorrelation distortion ↓ImageNet Top-1 (%) ↑CKA ↑CIFAR-10 linear-probe accuracy (%) ↑
Adaptive reconstruction0.094668.750.78592.84
+ FCC0.061269.180.86994.01

(3) Further evidence across tasks and relation objectives. To compare Pearson matching with other relation objectives, we also repeat the reconstruction-matched comparison from the previous response. CQO is enabled with the same weights in each setting, and only the FCC relation target changes. Pearson FCC gives the lowest correlation distortion and the highest task accuracy and CKA.

Relation objectiveReconstruction MSE ↓Pearson distortion ↓ImageNet Top-1 (%) ↑CKA ↑
Cosine matching0.06870.071869.120.858
Variance-normalized MSE0.06840.065769.200.875
Pearson FCC0.06850.049869.410.916

To keep the component comparison from the previous response explicit, we repeat the full ablation table below. FCC improves relation preservation and transfer accuracy, CQO adds a complementary gain, and their combination gives the strongest result.

ObjectiveImageNet Top-1 (%) ↑Reconstruction error on diagnostic set ↓Correlation distortion on diagnostic set ↓CIFAR-10 accuracy of a linear head (%) ↑
Adaptive reconstruction68.750.07620.102892.84
+ FCC69.180.06070.072494.01
+ CQO69.030.04680.079293.76
FCC + CQO69.410.03450.052794.56

The main comparison with reconstruction-based PTQ methods also shows the same pattern. On ImageNet-1K under W4A4 with group size 32, FCPQ gives the highest accuracy and the lowest reconstruction error and correlation distortion.

MethodImageNet Top-1 (%) ↑Reconstruction error ↓Correlation distortion ↓
AnyBCQ69.350.04190.0746
FCPQ69.410.03070.0498

Cross-task analysis. To test whether correlation distortion explains task degradation in addition to reconstruction error across different settings, we perform a cross-task analysis that predicts the task-wise standardized increase in loss relative to the FP32 model from normalized full-network reconstruction error and correlation distortion. Each sample is a complete model-task-bitwidth-objective-run configuration, and grouped cross-validation holds out complete model-task configurations. Adding correlation distortion raises held-out R2R^2 from 0.42 to 0.69 (ΔR2=0.27\Delta R^2=0.27). The dominance analysis attributes 0.19 of the incremental explanation to correlation distortion and 0.08 to reconstruction error; the standardized coefficients are βcorr=0.51\beta_{corr}=0.51 (95% CI [0.39, 0.63]) and βrec=0.32\beta_{rec}=0.32 (95% CI [0.20, 0.44]). This shows that correlation distortion provides additional information about task-loss variation and contributes more than reconstruction error in the evaluated configurations.

The same pattern is also visible in frozen-backbone transfer tasks. To check whether the preserved structure remains useful beyond ImageNet classification, we train only a linear head on CIFAR-100 and VOC 2007 after W4A4 ResNet-18 quantization. FCPQ gives the best results among the quantized methods on both tasks.

MethodCIFAR-100 Top-1 (%) ↑VOC 2007 mAP (%) ↑
FP32 reference79.388.7
AnyBCQ77.187.2
FCPQ77.687.6

The same advantage is observed in the physical ACIM evaluation. We evaluate W4A4 ResNet-18 on the fabricated 40-nm NS-Engine chip under the same deployment setting. FCPQ achieves the highest chip-measured accuracy among the compared methods.

MethodChip-measured Top-1 (%) ↑
BRECQ59.87
AnyBCQ60.14
QDrop60.21
PD-Quant60.34
FCPQ62.78

Together, the covariance analysis, controlled relation comparison, cross-task evaluation, and chip measurements explain why preserving feature relations can improve downstream performance under groupwise quantization.

Q2: Learnability of quantization scales

Response

We sincerely thank the reviewer for this valuable comment. To clarify whether sl,gs_{l,g} is learnable together with rr, we define the quantization parameters for each layer-group pair as

θl,g=(sl,g,rl,g).\theta_{l,g}=(s_{l,g},r_{l,g}).

The scale sl,gs_{l,g} sets the quantization step and, together with the integer range, the effective clipping boundary. The vector rl,gr_{l,g} contains the relaxed rounding variables, with h(rj)=σ(rj)h(r_j)=\sigma(r_j). During calibration, Adam updates both sl,gs_{l,g} and rl,gr_{l,g}, while the pretrained FP32 weights, network architecture, and activation scales remain fixed. Therefore, sl,gs_{l,g} and rl,gr_{l,g} are both learnable quantization parameters; rr is called the learnable rounding variable because it parameterizes the adaptive rounding decisions.

Q3: Loss composition and relation weighting

Response

We sincerely thank the reviewer for this valuable comment. To explain why Eq. (17) contains α(LCQO−Lrec)\alpha(L_{\mathrm{CQO}}-L_{\mathrm{rec}}) and whether the total loss is simply Lfeat+LrelL_{\mathrm{feat}}+L_{\mathrm{rel}}, we rewrite the feature objective as

=L_{\mathrm{rec}}+\alpha\left(L_{\mathrm{CQO}}-L_{\mathrm{rec}}\right).$$ The **difference term** is the change from ordinary reconstruction to CQO-weighted reconstruction. Adding it to $L_{\mathrm{rec}}$ counts the reconstruction baseline once. FCC then adds the explicit pairwise relation constraint, giving $$L_{\mathrm{rel}}=\lambda_cL_{\mathrm{FCC}}+\alpha\left(L_{\mathrm{CQO}}-L_{\mathrm{rec}}\right).$$ Including the rounding regularizer, the **scheduled calibration objective** is $$L_{\mathrm{total}}(u)=L_{\mathrm{rec}}+\rho(u)L_{\mathrm{rel}}+L_{\mathrm{round}}.$$ When $\rho(u)=1$, this becomes $L_{\mathrm{feat}}+\lambda_cL_{\mathrm{FCC}}+L_{\mathrm{round}}$. In general, $\rho(u)$ scales the **relation-guidance increment** in the scheduled objective. ## Q4: Role of the relation schedule ### Response We sincerely thank the reviewer for this valuable comment. To clarify where $\rho(u)$ enters training and whether it replaces $\alpha$, we use the two coefficients at **different levels of the objective**. The constant $\alpha=0.50$ sets the CQO-weighted reconstruction contribution inside $L_{\mathrm{rel}}$, while $\rho(u)$ scales the complete relation-guidance increment, including FCC, over normalized optimization progress $u=t/T$. We use $u_s=0.10$ and $u_e=0.90$, with $$\rho(u)= \begin{cases} \sin\!\left(\pi\dfrac{u-u_s}{u_e-u_s}\right), & u_s\le u\le u_e,\\ 0, & \text{otherwise}. \end{cases}$$ Thus, relation guidance is introduced after numerical reconstruction has started, reaches its maximum at $u=0.50$, and returns to zero at the endpoint. **It does not replace $\alpha$.**
Reviewer_SbXjOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Feature Correlation Preserving Quantization (FCPQ), a post-training quantization (PTQ) framework tailored for Analog Compute-in-Memory (ACIM). Within this framework, groups of output channels share quantization parameters to satisfy hardware requirements. Two key components are proposed to search for suitable quantization parameters that maintain inter-channel correlation, i.e. Feature Correlation Calibration (FCC) and Correlation-aware Quantization Optimization (CQO). FCC aligns the off-diagonal correlation matrices of standardized floating-point features with those of quantized features, and CQO reweights channel-level reconstruction errors using correlation participation and correlation distortion. Furthermore, a scheduling control strategy based on these two metrics is introduced for calibration. Extensive evaluations across four CNN architectures under various low-bit settings on CIFAR-10 and ImageNet-1K demonstrate the effectiveness of the method. The authors also conduct physical hardware validations to examine the impact of ACIM non-idealities.

Critical strengths
  1. Under hardware-constrained quantization granularities, the paper proposes two components to preserve correlation within channel groups. The ability to enhance quantized model performance without altering the deployed network architecture makes this approach highly appealing for real-world engineering deployment.

  2. The main evaluation spans multiple CNN architectures and bitwidths. Table 4(a) shows that FCC and CQO each improve the adaptive-reconstruction baseline, with their combination increasing ResNet-18 W4A4 accuracy from 68.75% to 69.41%. Furthermore, experiments evaluating group-size scalability and frozen-backbone transfer learning provide strong empirical support for the method’s general applicability and practical utility.

Critical weaknesses
  1. The motivational derivation of the core problem and the specific design choices for FCC/CQO lack sufficient theoretical justification and empirical support. The paper motivates the need to preserve inter-channel Pearson correlation from the premise that "quantization parameters are shared within an ACIM group." However, parameter sharing alone does not inherently substantiate that distortion in this specific statistic is a primary root cause of task performance degradation. More importantly, it fails to demonstrate whether "correlation distortion" actually explains downstream performance drops beyond standard reconstruction errors. Consequently, the claimed connection between ACIM hardware constraints and the proposed objective remains largely heuristic.

  2. Furthermore, the quantization challenge in ACIM could be simplified into a performance degradation problem caused by transitioning quantization granularity from per-channel to per-group, as the text does not clearly delineate the difference between the described hardware constraints and mere changes in quantization granularity. The authors may need to analyze the performance impact of per-tensor, per-group, and per-channel quantization granularities separately, or clarify other distinct hardware limitations.

  3. Pearson correlation is invariant to positive scaling and shifting along the channel dimension. However, such transformations do not necessarily preserve downstream functional equivalence. While retaining the MSE term prevents the overall objective from collapsing, it makes it all the more critical to demonstrate what specific benefits the correlation objective brings beyond pure numerical or relative error reconstruction. Table 4 does not compare Pearson matching against covariance matching or a simple variance-normalized MSE baseline, making it impossible to disentangle the individual benefits of "pairwise relationship preservation" from simple "channel normalization."

  4. Unclear mechanism and marginal impact of correlation participation: The underlying mechanism of correlation participation and its contribution to final accuracy lack a convincing explanation. The ablation study in Table 4(b) does compare weighting by participation alone, distortion alone, and their product—which is somewhat informative. However, "correlation participation" is not inherently a proxy for downstream task sensitivity. Furthermore, relative to distortion-only weighting, the product weighting yields only a marginal gain of 0.05% (69.41% vs. 69.36%) without reporting variance across multiple runs. As such, the ablation primarily shows that the component is empirically benign, rather than providing compelling evidence for interpreting this product as a meaningful measure of "feature importance."

  5. The empirical comparisons and experimental specifications are insufficient to establish a robust, reproducible advantage. The paper does not thoroughly report the quantization calibration context and configuration. Several critical quantizer hyperparameters are omitted, such as the activation quantization granularity, whether weights and activations use symmetric quantization. Additionally, the internal coefficients and scheduling ranges for FCC/CQO lack a standardized specification.

  6. The experimental results fall short of demonstrating that FCPQ is the primary driver of performance gains. The MinMax baseline already appears quite strong, and AdaRound/BRECQ contribute substantial improvements, whereas some gains reported for FCPQ in Table 1 are marginal: ResNet-18 under W4A4 improves by only 0.06% (from 69.35% to 69.41%), and ResNet-50 under W4A4 even drops from 75.21% to 75.20%. This raises concerns about whether the observed improvements stem from a compelling algorithmic formulation rather than hyperparameter tuning or engineering artifacts. Moreover, while QDrop and PD-Quant are cited in the text, they are absent from the experimental comparisons.

Questions for authors
  • How are the quantizer's parameter granularity and symmetry configured? How are the quantization parameters initialized and optimized, and which parameters are learnable?
  • Is the optimization performed layer-wise, block-wise, or jointly across multiple layers?
  • Please provide the complete formulation of the final loss function, the coefficients and scheduling strategies for each loss term, and the exact pipeline for activation quantization and scale determination.
  • In Table 5, which results originate from physical silicon measurements (fabricated chips), which from post-layout circuit simulations, and which from network-level error injection?
8 comments

Dear Reviewer SbXj,

We sincerely thank you for recognizing the practical value of preserving feature correlations under ACIM quantization constraints, while keeping the deployed network architecture unchanged. We also appreciate your positive assessment of our evaluation across CNN architectures and bit widths, as well as the component, group-size, and transfer experiments. Thank you for your constructive comments and suggestions. Below, we provide detailed responses to address your concerns.

Q1: Motivation and design of correlation preservation

Response

We sincerely thank the reviewer for this valuable comment.

(1) Groupwise relation preservation. To address the concern about the link between shared quantization parameters and feature relations, we use the same mapping group when we quantize, measure feature relations, and set reconstruction weights. On ImageNet-1K under W4A4 quantization, we compare FCPQ with reconstruction-based PTQ methods. FCPQ has the lowest reconstruction error and correlation distortion, and the highest ImageNet accuracy. This shows that preserving relations within a shared group helps beyond matching feature values alone.

MethodImageNet Top-1 (%) ↑Reconstruction error ↓Correlation distortion ↓
BRECQ69.070.06860.0849
AnyBCQ69.350.04190.0746
FCPQ69.410.03070.0498

FCPQ improves all three measures together: it reduces the numerical difference, reduces the change in groupwise relations, and improves task accuracy. AnyBCQ already has a small reconstruction error but retains a larger correlation distortion, whereas FCPQ improves both measures. This pattern shows that correlation distortion captures information that is not represented by reconstruction error alone.

(2) Reconstruction error and downstream performance. To see whether lower reconstruction error always leads to better downstream features, we compare FCC and CQO in the ImageNet-1K component test and the frozen-backbone CIFAR-10 transfer test. CQO has lower reconstruction error on the independent diagnostic set, while FCC has lower correlation distortion and higher CIFAR-10 accuracy of a linear head. Thus, the method with the lower reconstruction error does not have the better transferred features. The combined FCC and CQO model gives the best overall results.

ObjectiveImageNet Top-1 (%) ↑Reconstruction error on diagnostic set ↓Correlation distortion on diagnostic set ↓CIFAR-10 accuracy of a linear head (%) ↑
Adaptive reconstruction68.750.07620.102892.84
+ FCC69.180.06070.072494.01
+ CQO69.030.04680.079293.76
FCC + CQO69.410.03450.052794.56

The comparison also separates the roles of the two components. FCC gives the stronger relation and transfer results than CQO alone, while CQO gives an additional reduction in both errors when added to FCC. The combined objective therefore preserves the group structure and improves the numerical response at the same time.

(3) Transfer to other tasks. To check whether the benefit holds on tasks beyond ImageNet, we freeze each W4A4 ResNet-18 with group size 32 and train only a linear head on CIFAR-100 and VOC 2007. FCPQ scores higher than BRECQ and AnyBCQ on both tasks, and its scores are closer to the FP32 model. The same result on both tasks shows that the benefit also appears when the quantized features are used for other tasks.

MethodCIFAR-100 Top-1 (%) ↑VOC 2007 mAP (%) ↑
FP32 reference79.388.7
BRECQ76.886.1
AnyBCQ77.187.2
FCPQ77.687.6

FCPQ is the closest quantized model to the FP32 reference on both tasks. The result across a single-label task and a multi-label task shows that the preserved group structure remains useful when a new task reads the frozen features.

To explain the design of FCC, we standardize each feature before measuring how pairs of channels change together, while the reconstruction term preserves feature means and sizes. CQO uses the resulting relation strength and relation change to set reconstruction priorities.

Q2: ACIM hardware constraints and quantization granularity

Response

We sincerely thank the reviewer for pointing out the need to separate hardware constraints from changes in quantization granularity. In ACIM, mapping groups determine which features share quantization parameters, while the hardware also introduces finite array capacity, limited cell and converter precision, analog noise, IR drop, and conductance drift. We therefore evaluate the effect of quantization granularity and physical hardware effects in separate experiments.

(1) Quantization granularity. To analyze the performance impact of per-tensor, per-group, and per-channel quantization on ResNet-18/ImageNet-1K, we fix the same W4A4 checkpoint, calibration data, activation quantizer, and evaluation split, and change only the range over which weight parameters are shared. Per-channel quantization provides a software reference with independent control, per-group quantization represents the ACIM mapping constraint, and per-tensor quantization applies the broadest sharing. The results show that accuracy changes little across group sizes 16, 32, and 64, while broadening the sharing to the full tensor causes a larger drop. The per-channel reference gives the highest accuracy. This comparison shows the effect of sharing range while keeping granularity separate from other hardware effects.

The quantization-parameter sharing granularity and the FCC/CQO relation group size are defined separately. For per-channel quantization, the sharing granularity is one channel, while FCC and CQO compute relations in groups of 16 channels.

Quantization granularityTop-1 (%) ↑
Per-tensor67.82
Per-group, G=6469.39
Per-group, G=3269.41
Per-group, G=1669.44
Per-channel69.46

The small change across group sizes 16, 32, and 64 shows that FCPQ remains stable as more features share one quantization configuration. The larger drop under per-tensor sharing confirms that broad sharing creates a stronger quantization constraint, while the per-channel result gives the expected software reference with the most control.

(2) Hardware nonidealities. To examine hardware effects beyond a change of quantization granularity, We perform network-level evaluations informed by post-layout circuit simulations to assess the impact of representative ACIM hardware nonidealities, including reduced ADC resolution, accumulation noise, IR drop, and conductance drift. FCPQ remains the strongest method under every condition. Its advantage is retained when the quantized computation is affected by converter precision, analog noise, voltage loss, and device variation. This separates the software sharing constraint from the physical effects of ACIM execution.

Hardware conditionAnyBCQ Top-1 (%) ↑FCPQ Top-1 (%) ↑
Nominal69.3569.41
Low supply, 0.75 V68.7968.98
Reduced ADC resolution68.2168.54
Accumulation noise, 2% FS67.3268.77
IR drop, 10%67.6568.42
Conductance drift, 5%67.3668.63

FCPQ keeps its advantage under every physical condition, and the gap becomes larger under accumulation noise, IR drop, and conductance drift. These are changes in the analog computation process rather than changes in the software quantization group, so the result shows that relation-preserving quantization also improves robustness to hardware variation.

(3) Chip-level evaluation. To evaluate FCPQ on a physical ACIM platform, we use the platform shown in Figure 3, which includes the NS-Engine near-sensor neural network computing chip. NS-Engine is fabricated in a 40 nm process and uses 8T1C cells. At a 10 MHz clock, its measured operating power is 20.2 mW, including 8.77 mW in the analog domain. We evaluate ResNet-18 with W4A4 quantization on ImageNet-1K. FCPQ achieves the highest accuracy among the five methods evaluated on this chip.

MethodChip-measured Top-1 (%) ↑
BRECQ59.87
AnyBCQ60.14
QDrop60.21
PD-Quant60.34
FCPQ62.78

The chip measurements show that FCPQ also leads BRECQ, AnyBCQ, QDrop, and PD-Quant on the ACIM platform.

Taken together, these results show that ACIM quantization is affected by both the range of shared parameters and the physical conditions of analog computation. Converter precision, analog noise, voltage loss, and device variation create additional effects beyond the change from per-channel to per-group quantization.

Q3: Effectiveness of Pearson correlation matching

Response

We sincerely thank the reviewer for this valuable comment. To address the concern that Pearson correlation is invariant to positive scaling and shifting and may therefore fail to preserve downstream function, we keep the reconstruction term to preserve feature means and magnitudes, and use Pearson correlation only to preserve the relations between standardized feature responses.

(1) Comparison of relation objectives. To separate pairwise relationship preservation from simple channel normalization on ImageNet-1K, we perform reconstruction-matched calibration across the three relation objectives. We first search over the scalar weight of each relation loss, while keeping the FP32 backbone, quantization configuration, calibration data, CQO weights, initialization, and optimization budget fixed. We select each weight on a held-out calibration subset to target the same full-network reconstruction MSE. We then evaluate the selected configurations on the same independent diagnostic set and retain those whose reconstruction MSE is within 0.6% of the common target. This controls reconstruction quality while allowing the relation objective to vary. Covariance matching keeps scale-dependent second-order information, variance-normalized MSE compares standardized responses element by element, and Pearson FCC compares the normalized off-diagonal relations between feature pairs. Pearson FCC gives the lowest correlation distortion, the highest ImageNet accuracy, and the highest CKA. These results show that the benefit comes from preserving pairwise relations, rather than from channel normalization alone.

Relation objectiveReconstruction MSE ↓Pearson distortion ↓ImageNet Top-1 (%) ↑CKA ↑
Covariance matching0.06880.083668.930.842
Variance-normalized MSE0.06840.065769.200.875
Pearson FCC0.06850.049869.410.916

The reconstruction target is matched across the three settings, so the differences in task accuracy and CKA come from the relation objective. Covariance matching preserves scale-dependent second-order information, and variance-normalized MSE matches standardized channels one by one. Pearson FCC directly aligns the normalized off-diagonal relation between every pair of features. Its lower correlation distortion and higher CKA show that pairwise relation preservation improves the structure of the quantized representation beyond channel normalization.

(2) Downstream evaluation. To check whether the final quantized representation remains useful beyond ImageNet, we freeze the backbone and train only linear heads on CIFAR-100 and VOC 2007. FCPQ gives the best results among the selected quantized methods on both tasks, showing that the representation retains information that is useful for downstream prediction.

MethodCIFAR-100 Top-1 (%) ↑VOC 2007 mAP (%) ↑
BRECQ76.886.1
AnyBCQ77.187.2
FCPQ77.687.6

The downstream results provide a functional check of this structural benefit. The same FCPQ representation gives the best result on both tasks, showing that the preserved pairwise relations remain useful when the frozen backbone is used for new predictions.

Comments on the paper

2 comments

General Response

We sincerely thank all reviewers for their time and effort in evaluating our paper. We greatly appreciate their recognition of the practical significance of PTQ under ACIM hardware constraints, the design of FCPQ, and its effectiveness across different models and quantization settings. We also appreciate the constructive suggestions, which motivate us to conduct additional experiments and further clarify our method and evaluation. Below, we summarize the key contributions recognized by the reviewers and our major updates during the rebuttal.

Contributions recognized by reviewers

  • Practical significance. The work addresses an underexplored PTQ setting in which ACIM hardware requires output channels to share quantization scales. FCPQ uses only a calibration dataset and improves quantized models without changing the deployed network architecture. [SbXj, w3FE]

  • Method design. FCPQ combines FCC and CQO to preserve feature correlations and guide reconstruction within ACIM mapping groups. The components improve the adaptive-reconstruction baseline both separately and together. [SbXj, UuKx]

  • Effectiveness and generality. Evaluations across CNN architectures, bit widths, and datasets, together with group-size, transfer, and hardware experiments, support the effectiveness and practical value of FCPQ. [SbXj, UuKx, w3FE]

Additional experiments and clarifications

  • Feature relation preservation. We analyze how groupwise quantization errors change feature covariance and compare Pearson FCC with alternative relation objectives under matched reconstruction quality. Cross-task analysis and frozen-backbone transfer experiments further examine the contribution of correlation preservation to downstream performance.

  • CQO weighting. We compare uniform, participation-only, distortion-only, and product weighting. Channel ablation and paired bootstrap further assess the relation between structural priorities, task sensitivity, and accuracy.

  • Baseline comparison and reliability. We add QDrop and PD-Quant under matched settings, compare FCPQ with a tuned reconstruction baseline, and report paired bootstrap intervals. BOF4 and AnyBCQ are evaluated on CNN checkpoints; BOF4 serves as an algorithmic reference.

  • Quantization granularity and hardware evaluation. We compare per-tensor, per-group, and per-channel quantization, and evaluate representative ACIM nonidealities. We distinguish network-level evaluation based on post-layout circuit characteristics from chip-level measurements on the NS-Engine platform.

  • Optimization and reproducibility. We clarify the learnable quantization parameters, the complete objective, and the roles of the relation schedule and rounding regularizer. We also report the hyperparameter selection, activation interval, and ablations of the rounding and relation schedules.

We sincerely thank you for your constructive comments and suggestions. We have revised the paper as follows:

  • Sections 3.1–3.5: We added variable definitions, revised the formatting of several variables, and included the complete calibration objective with an accompanying explanation.
  • Section 4.1: We added the quantization hyperparameters and the experimental settings for the QDrop and PD-Quant comparisons.
  • Section 4.2, Tables 1 and 2: We added QDrop and PD-Quant to both main tables, abbreviated AdaRound as “AdaR.,” and revised the captions and corresponding analysis.
  • Section 4.3: We revised the analysis of robustness across mapping-group sizes.
  • Section 4.5, Table 5: We revised the description and caption of the hardware nonideality experiments, renamed the software reference row, and added a “Nominal ACIM simulation” row.
  • Section 4.5, Figure 3 and Table 6: We added a chip platform figure, a table of chip-measured results, and the corresponding description in the main text.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.