{"id":"da860780-0bbe-4bef-b36b-dd2611db12c4","arxiv_id":"2501.04517","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"HEQ sets quantization step sizes from weight quantiles, achieving competitive CIFAR-10 accuracy and enabling logic-gated residual networks on STL-10.","lead":"This paper introduces Histogram-Equalized Quantization (HEQ), which tunes the step size of low-precision neural network weights by aligning quantization thresholds with the data's own quantiles each epoch. The authors also design OR and MUX logic-gated residual connections, aiming to replace full-precision skip connections and lower the hardware cost of inference.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The CIFAR-10 state-of-the-art claim is based on a 0.01% margin without error bars, and the STL-10 comparison varies multiple factors at once, so the empirical support for the central claims is not established.","rationale":"The paper's central contribution is a simple, quantile-based step-size update rule and a logic-gated residual block. I find the algorithmic description clear and the method plausibly competitive; the use of a single parameter-free update is a useful idea. However, the advertised claims are empirical, and the evidence presented is not sufficient to support them. On CIFAR-10, HEQ-ternary outperforms LQ by only 0.01% at the same bitwidth, which is negligible without error bars. On STL-10, the comparison in Table II is confounded by differences in architecture, parameter count, bitwidth, and training method, so the 0.37% accuracy advantage cannot be attributed to the proposed components. A matched comparison with multiple seeds and a significance test is needed. The reader's weakest_assumption concerns the symmetry of the quantiles in Eq. (3); while that is a valid theoretical concern, it is not the most load-bearing because the update rule could still work as a heuristic even if the equalization is imperfect. The missing controlled experiments directly undermine the paper's core claims, so addressing them should be the priority.","tokens_in":8137,"tokens_out":18750,"duration_ms":166350,"concrete_test":"Re-run the CIFAR-10 comparison with fully matched conditions: same VGG-Small architecture, same pre-trained initialization, same augmentation, same optimizer, same 100-epoch learning-rate schedule, and the same batch size, for HEQ-ternary, LQ, and LLSQ at 2/2 bitwidth, over at least 10 seeds. Report mean±std and a paired Wilcoxon signed-rank test between HEQ and the best competitor; if the difference is not significant at p<0.05, the state-of-the-art claim is not supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Table I reports HEQ-ternary at 93.51% versus LQ at 93.50% for 2/2 bitwidth on CIFAR-10, a difference of 0.01 percentage points, with no error bars or significance tests despite claiming an average over 5 realizations. This margin is within typical run-to-run variance, so the claim that HEQ achieves state-of-the-art performance is not supported. Table II on STL-10 compares ORNet-11 (VGG-11, 3.14M params, 2/1 bitwidth) and MUXORNet-11 against VGG-7 (4.57M params, 2.5/8 bitwidth) from [18]; architecture, parameter count, bitwidth, and training procedure all change simultaneously, so the 0.37% improvement cannot be attributed to HEQ or the logic gates. No ablation isolates the contribution of HEQ versus the OR/MUX gating, nor compares against full-precision residual additions. Thus the paper's central empirical claims lack a controlled basis.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Histogram-Equalized Quantization (HEQ), an adaptive linear symmetric quantization method that updates the per-layer step size at the beginning of each epoch using n-quantiles of the proxy weight distribution, with the goal of equalizing the histogram of quantized values. The method is applied to ternary, quinary, and septenary weights with 2-bit activations on CIFAR-10, where the authors report 93.51% accuracy for HEQ-ternary and claim state-of-the-art performance. The paper also introduces OR-gated and MUX-gated residual blocks intended to replace full-precision skip connections, and reports STL-10 accuracies of 83.34% (VGG-11), 83.82% (ORNet-11), and 84.17% (MUXORNet-11). The central claims are that HEQ automatically adapts quantization thresholds without learned step sizes and that it enables effective training of logic-gated residual networks at lower hardware complexity.","tokens_in":8361,"tokens_out":4178,"duration_ms":40315,"significance":"If the empirical claims hold, HEQ would be an attractively simple replacement for learned step-size quantization, requiring no extra parameters and producing integer-only weights (0, ±1, ±0.5) that are directly hardware-friendly. The logic-gated residual network idea is also interesting because it could eliminate full-precision additions in ResNet-style topologies. The paper is clearly written and the algorithm is easy to reproduce. However, the evidence presented is not yet sufficient: the CIFAR-10 state-of-the-art claim rests on a 0.01 percentage point margin with no error bars, and the STL-10 comparison varies architecture, parameter count, bitwidth, and training procedure simultaneously. The symmetry assumption underlying Eq. (3) is not validated. The significance is therefore conditional on stronger empirical support and a more controlled experimental design.","major_comments":[{"comment":"The derivation of the step-size update is built on the assumption q_{-i} ≈ -q_i, i.e., that layer-wise weight quantiles are symmetric around zero. This assumption is stated but not validated. If a layer's weight distribution is skewed, the computed s will not equalize the quantization bins, and the method's core rationale is undermined. Please provide quantitative evidence of symmetry across layers and training stages (e.g., skewness of the proxy weight histograms, or a comparison of q_{-i} and -q_i), or report the performance degradation on a deliberately skewed distribution.","section":"Section III-A, Eq. (3)"},{"comment":"The claim that HEQ achieves state-of-the-art accuracy on CIFAR-10 is based on HEQ-ternary at 93.51% versus LQ at 93.50% for 2/2 bit width, a difference of 0.01 percentage points. The paper reports averages over 5 realizations but gives no standard deviations, confidence intervals, or significance tests. This margin is well within typical run-to-run variance for CIFAR-10 training, so the result does not demonstrate superiority. Please report the per-realization results with error bars and, if possible, a paired significance test, or soften the state-of-the-art claim to a claim of competitiveness.","section":"Table I, CIFAR-10"},{"comment":"The STL-10 comparison changes multiple factors at once: the reference method [18] uses a VGG-7 backbone (4.57M parameters) with 2.5/8 bit width, while the proposed models use a VGG-11 backbone (3.14M parameters) with 2/1 bit width. Architecture, parameter count, bit width, and training procedure all differ, so the 0.37% improvement of MUXORNet-11 over [18] cannot be attributed to HEQ or to the OR/MUX gates. A controlled comparison is needed, for example VGG-7 with HEQ versus [18] under matching bit widths, or VGG-11 with and without logic gates under otherwise identical settings.","section":"Table II, STL-10"},{"comment":"There is no ablation that isolates the contribution of HEQ from the contribution of the OR/MUX gating. In particular, the paper does not compare ORNet-11 or MUXORNet-11 against the same VGG-11 backbone with a standard full-precision residual addition, nor against a version trained with a learned step-size quantizer. Without such ablations, the claim that 'HEQ enables a proper training' of logic-gated residual networks is not supported. Please add these baselines.","section":"Section IV-B and Fig. 5"}],"minor_comments":[{"comment":"The first two subfigure captions both read 'Full-precision model.' Please distinguish them, for instance by indicating the layer index or the weight/activation type.","section":"Fig. 1"},{"comment":"The notation for the rounding/flooring operation in Eq. (1) is garbled in the text. Please define it explicitly, e.g., as \\lfloor w/s \\rfloor or nearest-integer rounding.","section":"Eq. (1)"},{"comment":"The instruction 'Find n-quantiles of layer l' is underspecified. Please state the quantile estimator used and whether the quantiles are computed on the proxy weights or on the quantized weights.","section":"Algorithm 1"},{"comment":"The sentence 'this can be enhanced by forcing the weight median to zero, which has not been applied' is a self-acknowledged limitation. Please either implement the median-centering or provide evidence that it has negligible impact.","section":"Section III-A"},{"comment":"The MUX threshold uses a fixed factor of 0.5. This is an additional free parameter; please report a sensitivity analysis or at least justify the choice.","section":"Section IV-A"},{"comment":"The paper motivates HEQ by an entropy-maximization argument but reports no entropy statistics. Reporting the entropy of quantized weights for HEQ versus baselines would substantiate the mechanism.","section":"Section III-B"},{"comment":"Please clarify why HEQ-septenary is marked '+' rather than '++' in the HW-Compatibility column, and why a 3-bit representation is used for 7 levels rather than 8.","section":"Table I"}],"recommendation":"major_revision","confidential_remarks":"The manuscript carries an IEEE copyright notice and a DOI for ISCAS 2022. If this is a submission to a journal, the editor should verify that it is not a duplicate publication or that appropriate disclosure has been made. The technical novelty is moderate, and the main concerns are experimental rather than conceptual, so I believe the paper could be made acceptable with substantial additional experiments."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe one thing to know: the quantile-based step-size rule in HEQ is a real idea, and the OR/MUX logic-gated residuals are a nice hardware-oriented trick. But the paper's empirical evidence is far too thin to support the “state-of-the-art” claim—93.51% vs 93.50% on CIFAR-10 is within run-to-run variance, and the STL-10 comparison changes architecture, model size, bitwidth, and training method all at once.\n\nWhat is genuinely new: Eq. (3) gives a simple closed-form update for the step size from the layer's weight quantiles, generalizing the mean/std rules that only work for 2–3 bit quantizers. The logic-gated skip connections (OR, MUX-OR) replace full-precision additions with bitwise operations, which could matter on FPGAs or ASICs. The paper is clearly written and gives enough detail to re-implement—the algorithm and the quantization function are specified.\n\nThe soft spots are real. First, the symmetry assumption q_{−i} ≈ −q_i is asserted, not validated; Fig. 3 even shows more −1 than +1 weights, suggesting asymmetric layer-wise distributions. The derivation is a heuristic—it matches the average of quantiles to thresholds, not an optimality condition. Second, the experimental support is weak: no error bars or significance tests despite averaging over 5 seeds, a 0.01% margin over LQ on CIFAR-10, and a non-matched STL-10 baseline. There is no ablation isolating HEQ from the gating, nor a comparison against full-precision residuals. Finally, the “state-of-the-art” framing overstates what a 0.01% margin means.\n\nThe paper is worth reading for the ideas, though. I'd send it to a referee if it came across my desk as a journal submission—the methods are novel enough to deserve a careful look—but the referee would likely require stronger experiments, matched baselines, and code release. For my own work, I'd cite the quantile step-size heuristic as a simple alternative, but not the empirical claims.\n\nRecommendation: engage with it, but treat the numbers with a grain of salt.","headline":"Neat quantile-based step-size rule and clever logic-gated residuals, but the empirical claims rest on a 0.01% margin and unmatched baselines.","tokens_in":8849,"tokens_out":3806,"would_cite":true,"duration_ms":36923,"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":"A simple quantile-based step size rule can match learned quantization and enable logic-gated residual networks.","keywords":["quantized neural networks","histogram equalization","quantization-aware training","linear symmetric quantization","ternary weights","logic-gated residual networks","skip connections","hardware-efficient inference"],"falsifier":"Train the same CIFAR-10 model with weights initialized from a distribution with a nonzero per-layer mean, measure the asymmetry $\\sum_i(|q_{-i}| - q_i)$ at every epoch, and check both the quantized-bin balance and final accuracy; if layers with strong asymmetry show unbalanced bins and degraded accuracy despite HEQ, the symmetry assumption is the bottleneck.","tokens_in":7960,"feed_emoji":"🧮","tokens_out":7454,"duration_ms":68060,"temperature":0.7,"pith_summary":"This paper introduces Histogram-Equalized Quantization (HEQ), a way to choose the step size of linear symmetric quantized neural networks by looking at the layer-wise histogram of full-precision proxy weights. Instead of learning the step size through backpropagation, HEQ finds the n-quantiles of the weight distribution and sets the step size so that the quantized values fall close to those quantiles, balancing how often each discrete value is used. The authors report that this rule matches or exceeds learned-step-size methods on CIFAR-10, reaching 93.51% accuracy with 2-bit weights and activations, and no loss compared to the full-precision baseline when using 5 or 7 weight levels. On STL-10, the same rule makes it possible to replace full-precision residual additions with OR and MUX logic gates, giving 84.17% accuracy with a smaller model than a budget-aware learned-step-size baseline. The practical payoff would be cheaper hardware: quantized weights are exactly 0, ±1, or ±0.5, so multiplications become logic operations and bit shifts.","feed_headline":"Quantile rule matches learned quantization at lower hardware cost","feed_subtitle":"The same update lets OR and MUX gates replace full-precision residual additions, improving STL-10 accuracy.","key_machinery":"The central object is the quantile-matched step size. At the start of each epoch, each layer's full-precision proxy weights are sorted into n-quantiles $q_{-i}, q_i$; assuming $q_{-i} \\approx -q_i$, the step size is set by $$s = \\frac{4 \\sum_{i=1}^{(n-1)/2}(|q_{-i}|+q_i)}{(n-1)^2},$$ which is exactly the value that makes the quantization thresholds $(2i-1)s/2$ coincide with those quantiles. This turns the histogram into the quantizer's calibration signal, replacing learned gradients or fixed norm factors, and it is the mechanism that makes the quantized weights land on 0, ±1, and ±0.5, values that map directly to logic gates and bit shifts in hardware.","core_discovery":"The central claim is that a proper quantizer should spread the quantized weights evenly across the available discrete levels, and that this equalization can be achieved without learned parameters. For each layer, HEQ computes the n-quantiles of the current proxy weight histogram, assumes the positive and negative quantiles are mirror images around zero, and updates the single step size s with Eq. (3) so that the quantization thresholds coincide with those quantiles. Because the update is recomputed once per epoch from the histogram, the step size tracks the evolving distribution during training. Empirically this yields balanced ternary and quinary weight histograms, CIFAR-10 accuracy of 93.51% for 2/2-bit W/A quantization and 93.75% for septenary weights, and, on STL-10, proper training of residual networks whose skip connections are implemented by OR and MUX gates rather than 32-bit additions, reaching 84.17% accuracy.","pith_inferences":["The rule is stated only for odd n and for weights; extending the same quantile matching to activations, or to even n, is a natural test that the paper does not run.","If the symmetry assumption fails, a simple fix would be to recenter each layer's weights before computing quantiles; that would let HEQ work on skewed distributions and is directly measurable.","The logic-gate result suggests that full-precision skip connections are not what makes residual training stable once the weight histogram is well calibrated; this could change how residual blocks are mapped to hardware.","One testable consequence is that HEQ's advantage over learned step sizes should grow with the bitwidth, since quantile matching becomes more informative as n increases."],"forward_implications":["On CIFAR-10 with VGG-Small, HEQ-ternary reaches 93.51% accuracy at 2-bit weights and 2-bit activations, on par with learned quantization methods while using no learned step size.","HEQ-quinary and HEQ-septenary reach 93.66% and 93.75% respectively, matching or slightly exceeding the 93.68% full-precision baseline.","On STL-10, HEQ-ternary with the OR-gated variant reaches 83.82% and with the MUX-OR variant 84.17%, beating a regularized learned-step-size baseline that uses more parameters.","The quantized weights are exactly 0, ±1, or ±0.5, so hardware can implement them with logic operations and bit shifts rather than full-precision multipliers and scale factors.","Replacing residual additions and attention-like multiplications by OR and MUX gates is compatible with HEQ, reducing the hardware cost of skip connections without sacrificing accuracy."],"supporting_citations":[{"why":"Supplies the TWN baseline and the fixed norm factor τ=0.7 that HEQ replaces.","marker":"[12]"},{"why":"The linear symmetric quantizer with a learned step size that HEQ compares against on CIFAR-10.","marker":"[11]"},{"why":"The learned step size quantization method behind the STL-10 baseline in [18].","marker":"[15]"},{"why":"Defines the fixed-threshold 2-bit activation quantization scheme used in all experiments.","marker":"[22]"},{"why":"Learned quantization method whose 2/2-bit CIFAR-10 result HEQ matches.","marker":"[10]"},{"why":"Provides the CIFAR-10 dataset for the state-of-the-art benchmark.","marker":"[23]"},{"why":"Provides the STL-10 dataset for the logic-gated residual network experiments.","marker":"[34]"},{"why":"Defines residual skip connections that the proposed OR gate replaces.","marker":"[26]"},{"why":"Source of the attention-like multiplication that the MUX branch replaces.","marker":"[30]"}],"fun_headline_variants":["Quantile step size matches learned quantization, cuts hardware","Histogram equalization replaces learned scaling, no training","No-learn quantizer: same accuracy, lower complexity","Quantile rule enables logic-gated nets with fewer gates"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that each layer's weight distribution is symmetric enough around zero that the positive and negative quantile pairs cancel in the step-size formula; if the histogram is skewed, the step size is miscalibrated and the intended equalization is not achieved.","fun_headline_variants_meta":{"raw":{"variants":["Quantile step size matches learned quantization, cuts hardware","Histogram equalization replaces learned scaling, no training","No-learn quantizer: same accuracy, lower complexity","Quantile rule enables logic-gated nets with fewer gates"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000909,"raw_usage":{"total_tokens":3855,"prompt_tokens":838,"completion_tokens":3017,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":454,"completion_tokens_details":{"reasoning_tokens":2953}},"tokens_in":454,"tokens_out":3017,"duration_ms":23954,"temperature":1.0,"reasoning_tokens":2953,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T21:31:18.324445+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train the same CIFAR-10 model with weights initialized from a distribution with a nonzero per-layer mean, measure the asymmetry $\\sum_i(|q_{-i}| - q_i)$ at every epoch, and check both the quantized-bin balance and final accuracy; if layers with strong asymmetry show unbalanced bins and degraded accuracy despite HEQ, the symmetry assumption is the bottleneck.","supporting_citations":[{"cited_title":"Learned step size quantization,","cited_arxiv_id":null,"evidence_quote":"The learned step size quantization method behind the STL-10 baseline in [18]."},{"cited_title":"An analysis of single-layer networks in unsupervised feature learning,","cited_arxiv_id":null,"evidence_quote":"Provides the STL-10 dataset for the logic-gated residual network experiments."},{"cited_title":"Residual attention network for image classification,","cited_arxiv_id":null,"evidence_quote":"Source of the attention-like multiplication that the MUX branch replaces."}],"review_version":1}