{"id":"268bb811-cb0e-493d-9886-4889b97ddc97","arxiv_id":"2508.21524","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"A binary-weight multi-bit-activation quantization method, using mean-standard-deviation weight binarization and a cubic activation approximator, improves CIM CNN accuracy on CIFAR-10 and ImageNet and favors 4-bit activations.","lead":"This paper proposes a quantization method for AI chips that compute inside memory, using one-bit weights and multi-bit intermediate values, and reports accuracy gains on CIFAR-10 and ImageNet. The work suggests that a middle ground between fully binary and fully multi-bit networks may be the most efficient design point for next-generation low-power accelerators.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Accuracy gains over prior CIM methods are read off published numbers rather than controlled re-runs, so the claimed improvement range is not established; the 4-bit optimality table is also absent from the manuscript.","rationale":"The reader's weakest assumption is exactly the uncontrolled baseline comparison, and I agree with it. The paper's algorithmic construction is internally coherent: moment matching is a legitimate weight-quantization rule, and the G(a) activation approximator is plausible. However, no code is provided, no training hyperparameters are given, and no baseline is re-run under a common protocol. The claimed accuracy range is therefore not distinguishable from training-recipe differences. Separately, the 4-bit optimality conclusion relies on a table that is missing from the submitted manuscript, so that part of the central claim is unverifiable from the text. These issues do not prove the method is wrong; they mean the evidence as submitted is insufficient to establish the headline numbers. The appropriate verdict remains CONDITIONAL, pending controlled comparisons and the missing table.","tokens_in":8533,"tokens_out":7003,"duration_ms":67909,"concrete_test":"Request the authors' training configuration and code, then re-run the four Table I baselines (Huang et al., Q-PIM, EGQ, XNOR-RRAM) under BWMA's exact training protocol on CIFAR-10, using the same epochs, batch size, optimizer, learning-rate schedule, augmentation, and random seeds. Recompute the accuracy deltas in Table I; if any baseline matches or exceeds BWMA, or if the minimum gap drops below the claimed 1.44%, the headline accuracy improvement is unsupported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim (abstract; Section IV-B, Table I) is that BWMA improves ResNet-18 accuracy by 1.44-5.46% on CIFAR-10 and 0.35-5.37% on ImageNet over Huang et al., Q-PIM, EGQ, and XNOR-RRAM. These deltas are differences between BWMA's runs and top-1 numbers quoted from other papers. Section IV-A lists only the GPU; it does not state training epochs, optimizer, learning-rate schedule, batch size, data augmentation, or weight decay for either BWMA or any baseline. Quantization-aware training is sensitive to all of these, so the quoted baselines could easily differ from BWMA by more than the claimed gains for reasons unrelated to quantization. No baseline is re-run under BWMA's protocol. Additionally, the second half of the central claim—that 4-bit activation is the optimal hardware-accuracy trade-off—is supported by a sentence in Section IV-B pointing to 'Table ??'; that table does not appear in the manuscript. The submitted evidence therefore does not yet distinguish the method's contribution from training-recipe differences, and the 4-bit optimality result cannot be checked.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes BWMA, a quantization scheme for compute-in-memory CNN accelerators in which weights are binarized per layer to two values set to the layer-wise mean minus/plus the standard deviation, and activations are quantized to multiple bits through a differentiable approximation of the uniform quantizer built from a quadratic Dirac approximation G(a). Training uses quantization-aware training with a modified straight-through estimator. The experiments compare ResNet-18 accuracy on CIFAR-10 and ImageNet with previous CIM-oriented quantized networks, test binary-specific architectures (BDenseNet, MeliusNet, ReActNet), and report DNN+NeuroSim hardware simulations for VGG-8 and ResNet-20 on SRAM, RRAM, and FeFET crossbars. The paper claims accuracy gains of 1.44%-5.46% on CIFAR-10 and 0.35%-5.37% on ImageNet over prior CIM methods, and claims that 4-bit activation quantization is the optimal hardware-accuracy balance.","tokens_in":8809,"tokens_out":3470,"duration_ms":33250,"significance":"If fully substantiated, BWMA would provide a simple, hardware-aware quantization recipe that improves accuracy over published CIM-oriented quantizers while retaining binary-weight hardware efficiency. The hardware simulation across multiple device types is a useful strength, and the quadratic Dirac approximation is computationally cheap and plausibly effective. However, the central quantitative claims currently rest on cross-paper baseline comparisons, an unresolved missing table, and under-specified algorithm components, so the significance of the results as presented cannot yet be assessed reliably.","major_comments":[{"comment":"The accuracy improvements of 1.44%-5.46% on CIFAR-10 and 0.35%-5.37% on ImageNet are computed as differences between BWMA's runs and top-1 numbers quoted from prior papers. Section IV-A specifies only the GPU used for training; it does not report epochs, optimizer, learning-rate schedule, batch size, data augmentation, weight decay, or any other training hyperparameters for BWMA or for the baselines. Quantization-aware training is sensitive to all of these choices, so the reported gains could be caused by training-recipe differences rather than by the proposed quantization method. The paper should either re-run the baselines under an identical protocol or provide a strong justification that the published numbers are directly comparable. As written, Table I does not establish the claimed improvement.","section":"Section IV-B, Table I"},{"comment":"The 'closed-form solution' for weight binarization is a definition rather than a derivation: setting wb1 = mu - sigma and wb2 = mu + sigma makes Eqs. (2) and (3) identities, so the first two moments are matched by construction and no objective is minimized. If the method is intended as a heuristic for distribution alignment, this should be stated explicitly. If it is claimed to be optimal under some criterion (e.g., KL divergence or quantization error), the objective and its derivation are missing. This distinction matters because the paper's contribution list describes the method as 'deriving closed-form solutions' that 'significantly improve' representational capability; that claim needs support, for example an ablation against fixed {+1,-1} weights and against previous adaptive binary sets.","section":"Section III-B, Eqs. (2)-(4)"},{"comment":"The differentiable activation quantizer is not fully specified. The paper writes the approximation as sa*G(a - ci) and states that sa and ci denote the scale and center of each interval, but it never gives equations relating sa and ci to amax, amin, b, or delta from Eq. (6), nor does it describe how the centers and scales are derived from the quantization grid. Without these definitions, the proposed approximation to Qmbits is not reproducible and the training procedure is incompletely specified. Please provide the explicit construction and, ideally, a small algorithm or pseudocode.","section":"Section III-B, activation quantization"},{"comment":"The central claim that 4-bit activation is the optimal hardware-accuracy balance is supported only by a sentence pointing to 'Table ??', which does not appear in the manuscript. That said, the claim is not checkable as submitted. Additionally, Table III ('Mamba and Mamba 2 models under noise') is not referenced anywhere in the text, appears unrelated to binary-weight CNNs, and contains unexplained columns (e.g., '1.00 / 1.00' relative values). This table should be either removed or fully integrated and explained; in its current form it is a serious presentation and integrity issue.","section":"Section IV-B, Tables"},{"comment":"The modified straight-through estimator contains two free hyperparameters, temperature t and scaling factor alpha, but the paper does not report the values used in the experiments or any sensitivity analysis. Since this estimator directly controls the gradient approximation for binary weights, the absence of these values makes the training setup incomplete and prevents reproduction. Please report the chosen values and, ideally, an ablation showing the impact of t and alpha.","section":"Section III-B, Eq. (5)"}],"minor_comments":[{"comment":"There is a typo in the Index Terms: 'Model Qquantization' should read 'Model Quantization'.","section":"Index Terms"},{"comment":"The third contribution bullet states '0.35-5.46% accuracy improvements' without splitting the range by dataset, which conflicts with the abstract's '1.44%-5.46%' (CIFAR-10) and '0.35%-5.37%' (ImageNet). Please make the ranges consistent and clearly attribute them to datasets.","section":"Section I, contribution bullet"},{"comment":"The labels in Figure 2(b) are inconsistent: the left side uses 'sa * sign(a)' while the right side uses 'sa * G(a)'. The text says the differentiable function is sa*G(a - ci), so the figure should be aligned with the text to avoid confusion.","section":"Figure 2(b)"}],"recommendation":"major_revision","confidential_remarks":"The unresolved 'Table ??' reference and the unrelated Table III suggest the manuscript may be a draft with leftover content from another project. I recommend the editor verify that the submitted version is the intended one before further review, since these issues are more than stylistic and affect the verifiability of the main claims."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The one thing worth knowing: this paper has a clean, simple idea—binarize each layer's weights to mean ± standard deviation and quantize activations with a cubic straight-through estimator—and it evaluates the combination on real CIM hardware simulators. That is a plausible middle-ground recipe for CIM designers, and the hardware results (DNN+NeuroSim, multiple device types and crossbar sizes) are a genuine effort. But the evidence as submitted does not back the headline accuracy gains, and one of the two central claims is literally uncheckable.\n\nWhat's new and what's good: the weight rule (Eq. 4) is easy to implement and parameter-free, which is appealing in practice. The activation approximator (Eq. 8) is a reasonable STE variant. The hardware simulation section shows real work and identifies ADC/peripheral costs as the bottleneck, which is a useful observation. Credit where due: the paper is self-consistent and the math is straightforward.\n\nNow the soft spots, in order of severity. First, the accuracy improvements in Table I (1.44–5.46% on CIFAR-10, 0.35–5.37% on ImageNet) are differences against top-1 numbers quoted from other papers, not re-run baselines. The training recipe is under-specified (only GPU is given; no epochs, optimizer, schedule, augmentation, or weight decay). Quantization-aware training is sensitive to all of those, so the claimed gains could easily be training-recipe artifacts. Second, the 4-bit-optimality claim points to 'Table ??', which does not appear anywhere in the manuscript. That is a missing table, not a minor typo—it removes the basis for half the conclusion. Third, Table III compares Mamba and Mamba2 under noise; it has nothing to do with this paper and looks like an editorial mistake. That kind of artifact makes me worry about overall quality control. Fourth, no code is provided, which matters because the method's value is empirical.\n\nOn the weight derivation: calling it a 'closed-form solution' oversells it. Setting w_b to mean ± std guarantees moment matching by construction, but it is not derived from an optimality objective. It is a special case of AdaBin's center/distance parameterization, as the reader notes. That limits novelty but doesn't kill the contribution; the CIM context and hardware evaluation are still useful.\n\nWho this is for: people designing CIM accelerators who want a quick binary-weight method to try. It deserves a serious referee, but only after the authors supply the missing table, re-run or honestly qualify the baselines, and remove the Mamba artifact. I would not cite it in its current form.","headline":"Simple moment-matching binary-weight recipe with a cubic STE, but the accuracy claims rely on cross-paper numbers and the 4-bit optimality table is missing.","tokens_in":9273,"tokens_out":1492,"would_cite":false,"duration_ms":16884,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Binary weights set per layer to the mean plus or minus the standard deviation, paired with a differentiable 4-bit activation quantizer, raise accuracy over prior compute-in-memory quantization methods on CIFAR-10 and ImageNet.","keywords":["compute-in-memory accelerators","neural network quantization","binary weight networks","multi-bit activations","moment matching","differentiable quantization","hardware-accuracy trade-off","crossbar arrays"],"falsifier":"Re-run the Table I baselines on ResNet-18 under BWMA's exact training pipeline, using 1-bit weights and 4-bit activations for BWMA and the published settings for each baseline; if BWMA does not exceed each baseline on both CIFAR-10 and ImageNet, the reported 1.44-5.46 and 0.35-5.37 percentage-point gains do not hold. Separately, sweep activation bitwidths from 3 to 6 in the paper's hardware simulation and compute accuracy per unit energy and latency; if 5-bit activations dominate 4-bit on both metrics, the claim that 4-bit is optimal fails.","tokens_in":8331,"feed_emoji":"⚡","tokens_out":18151,"duration_ms":146575,"temperature":0.7,"pith_summary":"Compute-in-memory (CIM) accelerators promise large energy savings for neural networks, but they force weights and activations into coarse formats that usually cost accuracy. This paper tries to break the usual trade-off by using binary weights that are not fixed to $\\pm 1$: each layer chooses its own two weight values as $\\mu - \\sigma$ and $\\mu + \\sigma$, the mean and standard deviation of that layer's full-precision weights, so the binarized layer keeps the same first two moments. Activations, meanwhile, are quantized to a few bits with a purpose-built differentiable function that approximates ideal uniform quantization without searching for hyperparameters. On ResNet-18, the resulting 1-bit-weight, 4-bit-activation models reach 89.91% on CIFAR-10 and 67.85% on ImageNet, improving on prior CIM-oriented quantizers by 1.44-5.46 and 0.35-5.37 percentage points respectively. Hardware simulations point to 4-bit activations as the best balance between cost and accuracy across SRAM, RRAM, and FeFET accelerators.","feed_headline":"Binary weights + 4-bit activations lift CIM accuracy up to 5.5%","feed_subtitle":"Binary-weight matching plus a differentiable quantizer keeps 1-bit models accurate on CIFAR-10 and ImageNet.","key_machinery":"The load-bearing object is the layer-wise moment-matched binarization rule. For each layer's full-precision weights $w_i$, the two binary values are $w_{b1} = \\mu - \\sigma$ and $w_{b2} = \\mu + \\sigma$, where $\\mu$ and $\\sigma$ are the weight mean and standard deviation; this closed form forces the binarized distribution to share the original's first and second moments and gives each layer an adaptive binary set rather than a global $\\pm 1$. The second mechanism is a differentiable surrogate for uniform activation quantization: a quadratic derivative $g(a) = -2a^2 + 5/3$ on $|a| \\le 1$ integrates to $G(a) = -2a^3/3 + 5a/3$, whose scaled and shifted versions approximate the ideal multi-bit quantizer smoothly enough for backpropagation. Together these mechanisms let a CIM network train end-to-end with binary weights and multi-bit activations while keeping memory cells 1-bit and letting ADC/DAC resolution set the activation precision.","core_discovery":"On its own terms, the paper's central claim is that the accuracy penalty of binary weights in CIM accelerators can be largely erased by replacing the fixed $\\pm 1$ binarization with layer-specific values derived in closed form, while the activation side benefits from a differentiable approximation to uniform multi-bit quantization instead of a search-based quantizer. Concretely, quantizing weights to the per-layer pair $\\mu - \\sigma$ and $\\mu + \\sigma$ keeps the mean and variance of the original distribution, which the authors argue carries more information than a global two-value set; and the function $G(a) = -2a^3/3 + 5a/3$ on $[-1,1]$, clamped outside, gives a smooth surrogate for the sign function whose scaled and shifted copies approximate $b$-bit uniform quantization. With 1-bit weights and 4-bit activations, ResNet-18 reaches 89.91% on CIFAR-10 and 67.85% on ImageNet, which is 1.44-5.46 and 0.35-5.37 percentage points above four published CIM-oriented quantization baselines. The paper further claims that this configuration also generalizes to binary-specific architectures, adding 0.3-0.9 percentage points over BDenseNet, MeliusNet, and ReActNet-A, and that hardware simulations identify 4-bit data converters as the best accuracy-cost operating point across SRAM, RRAM, and FeFET crossbars.","pith_inferences":["The moment-matching binarization rule is not tied to CIM hardware, so the same closed-form weight rule could be tested in any quantization-aware training pipeline that wants binary weights with per-layer scaling.","The differentiable activation quantizer is a generic straight-through-estimator replacement and could be evaluated in other bitwidths and network families, potentially removing the need for clipping-range tuning in PACT-style methods.","The 4-bit-optimality result implies a concrete design rule for CIM chips: spend the hardware budget on ADC/DAC resolution rather than multi-bit memory cells, since the array itself contributes little to area and energy.","Editorial note: the paper includes an unreferenced third table that reports noise-robustness numbers for two models unrelated to the CNN/CIM experiments, so that table's evidentiary role is unexplained."],"forward_implications":["Binary weights can carry more information than a fixed $\\pm 1$ set: per-layer $\\mu \\pm \\sigma$ values improve accuracy on binary-specific networks by 0.3-0.9 percentage points over their original counterparts.","A 1-bit-weight, 4-bit-activation ResNet-18 outperforms published CIM-oriented quantizers with mixed-precision weights and activations, including EGQ's reported 6.1-bit weight and 6.3-bit activation setting, on both datasets.","Activation bitwidth, not the weight cell, is the main hardware lever: ADC and peripheral circuits dominate latency and energy, so designs should keep cells binary and choose converter resolution carefully.","Four-bit activations emerge as the cost-accuracy optimum across SRAM, RRAM, and FeFET devices and both tested crossbar sizes, giving a concrete design target for CIM accelerators.","Increasing crossbar size does not always reduce energy because unused-cell occupancy grows, so resource utilization must be traded against parallelism."],"supporting_citations":[{"why":"It supplies the mixed-precision baseline (6.1-bit weights, 6.3-bit activations) that BWMA reports beating by 5.46 points on CIFAR-10 and 5.37 points on ImageNet.","marker":"[10]"},{"why":"It supplies the flexible mixed-precision baseline with 5.2-bit weights and 8-bit activations against which BWMA reports a 3.68-point CIFAR-10 gain.","marker":"[11]"},{"why":"It supplies the strongest prior ImageNet baseline in the table (67.50%), which sets the 0.35-point margin BWMA claims to improve.","marker":"[12]"},{"why":"It supplies the binary-weight, binary-activation CIM baseline (88.47% on CIFAR-10) that the binarized-weight design must outperform.","marker":"[13]"},{"why":"It supplies the hardware simulation environment from which the latency, area, energy, and 4-bit-optimality findings are drawn.","marker":"[7]"},{"why":"It introduces adaptive binary sets for weights, the prior idea against which the paper's closed-form moment-matching binarization is positioned.","marker":"[20]"}],"fun_headline_variants":["Binary weights, 4-bit activations: CIM CNNs gain up to 5.5%","BWMA: 1-bit weights, multi-bit acts, 5.5% better CIM CNNs","Closed-form binary weights + differentiable quantizer lift CIM accuracy","Per-layer binary weights and 4-bit activations lift CIM CNN accuracy"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The accuracy gains are measured against published accuracy numbers from previous methods, not against those methods re-trained under the same schedule, augmentation, and hyperparameters as BWMA; if those baselines are not directly comparable, the claimed improvement margins are unsupported.","fun_headline_variants_meta":{"raw":{"variants":["Binary weights, 4-bit activations: CIM CNNs gain up to 5.5%","BWMA: 1-bit weights, multi-bit acts, 5.5% better CIM CNNs","Closed-form binary weights + differentiable quantizer lift CIM accuracy","Per-layer binary weights and 4-bit activations lift CIM CNN accuracy"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001097,"raw_usage":{"total_tokens":4646,"prompt_tokens":1081,"completion_tokens":3565,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":697,"completion_tokens_details":{"reasoning_tokens":3472}},"tokens_in":697,"tokens_out":3565,"duration_ms":25148,"temperature":1.0,"reasoning_tokens":3472,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T16:41:27.486437+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the Table I baselines on ResNet-18 under BWMA's exact training pipeline, using 1-bit weights and 4-bit activations for BWMA and the published settings for each baseline; if BWMA does not exceed each baseline on both CIFAR-10 and ImageNet, the reported 1.44-5.46 and 0.35-5.37 percentage-point gains do not hold. Separately, sweep activation bitwidths from 3 to 6 in the paper's hardware simulation and compute accuracy per unit energy and latency; if 5-bit activations dominate 4-bit on both metrics, the claim that 4-bit is optimal fails.","supporting_citations":[{"cited_title":"Mixed Precision Quantization for ReRAM-based DNN Inference Accelerators,","cited_arxiv_id":null,"evidence_quote":"It supplies the mixed-precision baseline (6.1-bit weights, 6.3-bit activations) that BWMA reports beating by 5.46 points on CIFAR-10 and 5.37 points on ImageNet."},{"cited_title":"Q-PIM: A Genetic Algorithm Based Flexible DNN Quantization Method and Application to Processing-in-Memory Plat- form,","cited_arxiv_id":null,"evidence_quote":"It supplies the flexible mixed-precision baseline with 5.2-bit weights and 8-bit activations against which BWMA reports a 3.68-point CIFAR-10 gain."},{"cited_title":"Genetic Algorithm-Based Energy-Aware CNN Quan- tization for Processing-In-Memory Architecture,","cited_arxiv_id":null,"evidence_quote":"It supplies the strongest prior ImageNet baseline in the table (67.50%), which sets the 0.35-point margin BWMA claims to improve."},{"cited_title":"XNOR-RRAM: A Scalable and Parallel Resistive Synaptic Architecture for Binary Neural Networks,","cited_arxiv_id":null,"evidence_quote":"It supplies the binary-weight, binary-activation CIM baseline (88.47% on CIFAR-10) that the binarized-weight design must outperform."},{"cited_title":"DNN+NeuroSim V2.0: An End-to-End Benchmarking Framework for Compute-in-Memory Accelerators for On-Chip Train- ing,","cited_arxiv_id":null,"evidence_quote":"It supplies the hardware simulation environment from which the latency, area, energy, and 4-bit-optimality findings are drawn."},{"cited_title":"Adabin: Improving Binary Neural Networks with Adaptive Binary Sets,","cited_arxiv_id":null,"evidence_quote":"It introduces adaptive binary sets for weights, the prior idea against which the paper's closed-form moment-matching binarization is positioned."}],"review_version":1}