acceptodds
Under review as a conference paper at ICLR 2027

Task-Aware QUBO Allocation for Mixed-Precision Quantization

Abstract

Mixed-precision quantization requires discrete allocation of weight and activation bit-widths, followed by recovery of the selected network. We develop a task-aware quadratic unconstrained binary optimization (QUBO) surrogate with separate weight and activation profiles, a bit-operation (BOP) cost, and selected structural priors, then refine the resulting route through direct validation-based PROTES search. Across three public pretrained SIDD checkpoints with source-image-disjoint calibration and evaluation, the production simulated-annealing QUBO routes are within 0.01 dB of the strongest SFC routes in a retrospective 19-solver comparison. Refinement gives single-run static gains of 0.27–1.01 dB over QUBO and 0.39–1.34 dB over the HAWQ-style allocator at approximately 4.1% routed-layer BOPs, while remaining 0.39–0.87 dB below FP32. On a compact NAFBlock-based denoiser evaluated on a held-out patch split with shared source images, the refined route improves static PSNR over HAWQ-style by 0.21–2.46 dB at nearby costs across the three tightest requested targets. Recovery narrows these differences and can change route ordering, so we evaluate allocation quality separately before and after LSQ+. We report achieved cost, search variation, and optimization expense; BOP reductions are analytical, and the reference inference path remains floating point.

Primary area:
optimization
Keywords:
mixed-precision quantization, neural network compression, QUBO, combinatorial optimization, PROTES, image denoising

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_y5GFOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

The paper introduces an approach for mixed-precision quantization. It formulates the problem as a quadratic unconstrained binary optimization problem, the solution of which is used as initialization for a PROTES refinement.

Critical strengths

The results show some improvements over the home-brewed baseline.

The authors provided several ablation studies for their method.

Critical weaknesses
  1. Clarity

1.1. The paper is in general hard to follow, using terms that have not been clearly defined (e.g., the notion of recovery, convolution route, menu). Importantly, the main motivation and contribution(s) are not clearly stated. I strongly suggest the authors to explain what are the limitations of the existing solutions and what they propose to address these limitations.

1.2. Some parts of the paper are explained in vague terms. For example, the different terms in Eqs. 1 and 2 are only defined in words. Some of them are more precisely defined in the appendix, but not all. As such, the paper would not be reproducible. I understand that reproducibility may be palliated by making the code public, but this would still leave the reader with a feeling of vague understanding of the work.

  1. Scope

2.1. Based on the beginning of Section 3 and on the experiments, the work seems to focus on convolutional neural networks. It is unclear to me whether this is a true restriction or if the framework would in general apply to other models, e.g., transformers. If it would, demonstrating this empirically would be needed.

  1. Experiments

3.1. The main baseline seems outdated and is a home-brewed version of an actual method HAWQ-V3, which dates back from 2021. It is unclear to me why not compare to the actual HAWQ-V3 method, and, more importantly, why not consider more recent baselines. For example, the second part of the related work section mentions more recent work, but it is unclear why they would not be applicable here.

3.2. The results focus on one model, thus casting doubts on the generalization of the approach.

Questions for authors

See Weaknesses above.

2 comments

Thank you for the careful reading. We answer each point below.

1.1 Terms, motivation and contribution. The revision will define every term at first use: "route" becomes bit-width assignment, "menu" becomes candidate bit-width set, and recovery is LSQ+ fine-tuning. Its introduction will state the gap directly. The methods discussed here make different trade-offs: HAWQ-V3 ties weight and activation precision, OMPQ, EMQ and CLADO fix activation precision, and EdMIPS updates network weights during its search. We allocate weight and activation precision per layer, jointly, under a hard BOP budget, without weight updates.

1.2 Equations and reproducibility. Every term of Eqs. 1–2 will have a formula in the main text and an entry in a notation table, and we will release the code with the final version.

2.1 Scope. It is not a restriction in principle, because the formulation extends to other operators given their damage and cost profiles. However, our pipeline currently quantizes convolutional layers only, not fully connected layers, and does not model the attention products between two activations, which have no learned weight operand and would need their own damage profiles and cost terms. We have not tested a transformer, so the revision will scope its claims to convolutional networks. To test transfer beyond denoising we added ResNet-20 on CIFAR-10. The candidate set transfers, but the weight/activation balance does not: with the denoiser's balance our allocation is below uniform, and with a balance selected on validation it is within 0.2 points of uniform, not better. The revision will report this as a limitation.

3.1 Baselines.

HAWQ-V3 as published. It gives each layer one bit-width shared by weights and activations (INT4 or INT8), so on our problem any budget below W8A8 (6.25 % BOPs) forces 4-bit activations. We tested this on HalfUNet with a HAWQ-style sensitivity, an exact solver, our quantizers and the skip constraint: before fine-tuning, test PSNR is 4.9 to 6.3 dB below uniform at our four budgets (2.9 to 4.3 dB with our own objective). This describes our static stack, not the published system, because we use dynamic activation quantization and did not fine-tune these rows, whereas HAWQ-V3 uses quantization-aware training. We therefore kept HAWQ-V3's Hessian-trace sensitivity and budgeted allocation but applied them to our joint W/A candidate set. The submitted row will be called "Hessian-trace knapsack", and a new row adds exact hard-budget allocation and a validation-tuned W/A balance.

More recent work. The QUBO papers in our related-work section address other compression problems (rounding at a fixed precision, joint pruning and quantization, channel pruning), so we cite them as related formulations and did not adapt them into bit-width allocators. The recent mixed-precision allocators OMPQ, EMQ and CLADO allocate weight bits only, under a model-size budget, so comparing them in our joint W/A space would require adding per-layer activation decisions and replacing the model-size constraint with a BOP budget.

What we added instead. Three baselines in our joint W/A space, each with three LSQ+ seeds: the tuned Hessian-trace allocator, a FIT-based allocator (FIT's metric, minimized with our exact solver) and EdMIPS's differentiable joint W/A search, ported to our candidate set and quantizers. After fine-tuning, their means and ours (annealed, corrected estimator) are within about 0.1 dB at every budget, while ours beats uniform and the submitted baseline in every seed at the three tight budgets (+0.13 to +0.43 dB over uniform). We do not claim better sensitivity terms: before fine-tuning the FIT-based allocator leads ours by +0.02 to +0.31 dB with the same tuning, and after fine-tuning EdMIPS is +0.02 to +0.08 dB ahead of ours at three of four budgets. EdMIPS, however, updates network weights and needed ten GPU searches (30 epochs each, our cap; none met our convergence criterion). Our contribution is therefore the joint, BOP-constrained formulation, which an exact MILP solves in under 2 s.

3.2 "Results focus on one model". The paper evaluates four denoisers: HalfUNet and three public checkpoints (NAFNet-32, Baseline-32, HINet-0.5×). On the public checkpoints, calibration and evaluation use disjoint source images, and with the corrected weight-error estimator (see our reply to Reviewer 2nLz) our allocation leads the tuned Hessian-trace allocator on all three before fine-tuning (+0.05 to +0.21 dB). The primary HalfUNet split does share source images between its partitions. The revision will state this, and a rerun on a source-image-disjoint split is scheduled, as is ResNet-20 fine-tuning.

AuthorsOpenReview ↗

Revised manuscript: clarity, baselines, and scope

1.1–1.2 Presentation and equations. Sections 1–3 now state the problem and contributions and define “bit-width assignment,” “candidate bit-width set” and “LSQ+ fine-tuning,” replacing “route,” “menu” and “recovery,” respectively. Eqs. 1–9 give formulas for the allocation terms in the original Eqs. 1–2, and Appendices A–C give notation and protocols. The general comment explains the estimator correction and code release.

3.1 Baselines. Table 2 reports similar means after LSQ+ for Hessian-trace (tuned), FIT-based (tuned) and EdMIPS (ported) and lists Hessian-trace knapsack (as submitted) separately (Section 5.1). The EdMIPS port reached its search cap without meeting our assignment stability criterion. Section 2 discusses LIMPQ (2022), the closest related integer-programming allocator, which requires quantization-aware training and was outside our experimental comparison. OMPQ, EMQ and CLADO allocate only weight widths, under model-size limits.

In our ablations, fixing activations at 8 bits is infeasible below 3.125% within the tested weight range. Relative to joint allocation by Ours (exact), it scores 0.57 dB lower before fine-tuning at 3.42% but 0.01–0.02 dB higher after fine-tuning at 3.42–4.10% (Table 21). Tested with our quantizers, HAWQ-V3’s tied W = A ∈ {4,8} space yields static PSNR 4.9–6.3 dB below same-budget uniform with Hessian-trace sensitivity and 2.9–4.3 dB below with our objective (Table 20). The test covers the tied search space alone.

2.1 Transformers. In the static comparison in Section 5.2 (Table 3), Restormer reaches 39.39 dB with Ours (annealed), 38.66 dB with uniform W6A7 and 39.17 dB with Hessian-trace (tuned). FIT-based (tuned) and Ours (exact) were not run. We allocate only convolutions and keep attention products, normalization and softmax FP32, so the 4.10% allocated-layer budget corresponds to 12.86% of FP32 BOPs for counted convolutions and attention products (Appendix C.1).

3.2 More models. Section 5.2 and Table 3 compare five networks before fine-tuning. On the source-disjoint split, which separates source images but can share scenes, Ours (exact) gains 0.57 dB over the best fitting uniform at 4.10% in all three seeds, while our variants change rank across fine-tuning seeds at 2.93% (Section 5.3, Figure 3). ResNet-20 shows no consistent fine-tuned gain over uniform (Section 5.5, Table 14).

Correction to 3.1. Our first reply’s “+0.13 to +0.43 dB” compared Ours (annealed) with the uniform configuration at each budget. Against the best fitting uniform, Ours (exact) gains 0.17 and 0.13 dB at 2.93 and 3.42% in all three seeds, with mean gains below 0.01 dB at 4.10 and 4.69% (Table 2 and Appendix D, Table 8).

Reviewer_M7Y9OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

In this work, authors introduce a mixed-precision quantization framework that selects the bit-widths for weights and activations to balance model accuracy and compression. Firstly, It uses a task-aware QUBO model to estimate suitable bit-width allocations. Secondly, it refines these allocations through direct validation using PROTES search. The QUBO model considers weight and activation precision separately, computational costs measured by Bit-Operations, and structural constraints. Its results are comparable to those of leading space-filling curve optimization methods. Experiments on SIDD benchmarks using NAFBlock-based denoising networks demostrate that PROTES refinement achieves substantial improvements in PSNR over standard QUBO and HAWQ-style methods under strict compression constraints. Finally, the framework evaluates quantization quality both before and after fine-tuning with LSQ+, explains changes in the ranking of bit-width allocations caused by fine-tuning.

Critical strengths

(1) The work combines a task-aware Quadratic Unconstrained Binary Optimization surrogate with a subsequent direct, validation-based tensor-train search. This setting allows the framework to bridge the gap between global surrogate estimation and direct task-evaluation refinement.

(2) Comparing with conventional layer-independent compression models, the proposed QUBO approximator incorporates separate weight-sensitivity and empirical global activation-sensitivity profiles. Moreover, It incorporates structural priors to better model real hardware/network behavior.

(3) The paper explicitly evaluates allocation quality both statically and post-recovery, showing that recovery fine-tuning can narrow initial gap advantages and alter route ordering.

Critical weaknesses

(1) Running direct refinement (PROTES) adds hundreds of seconds of compute time compared to pure proxy optimization. More over, end-to-end optimization cost rises from ~198 seconds for QUBO to ~734 seconds for QUBO+PROTES on the main model, and up to several hours on larger public checkpoints, which is quite long.

(2) In the work, terms like module-order interaction and skip-connection penalties are hand crafted structural heuristics rather than exact, graph-derived data-flow formulations. In addition, to prevent catastrophic routing failures ("activation cliffs") under QUBO search, the activation search space had to be manually restricted to {6, 7, 8} bits, which is too strict.

Questions for authors

(a) Since LSQ+ fine-tuning significantly narrows the static PSNR improvement of PROTES-refined routes over the pure QUBO (e.g., at the 2.95% BOP target), my question is : how would you recommend practitioners to decide whether the offline search cost of PROTES is reasonable when fine-tuning is part of the deployment pipeline?

(b) The activation candidate menu was manually limited to {6, 7, 8} bits to prevent activation cliffs in the process of QUBO search. Have you tested whether this specific menu boundary holds across non-denoiser vision architectures, or will new domains also require re-profiling this search space?

3 comments

Thank you for the positive assessment and the practical questions.

W1 / (a) — Is the PROTES cost worth it when fine-tuning follows? At the tight budgets we tested, no, and the revision will say so. Refinement helps before fine-tuning: it adds +0.19 to +0.71 dB test PSNR on HalfUNet (with the corrected weight-error estimator described in our reply to Reviewer 2nLz). After LSQ+ on HalfUNet, however, the difference over three seeds is only −0.07 to +0.06 dB. A search takes about 490–500 s on one H200, the time of 47 extra fine-tuning epochs, and at 2.93, 3.42 and 4.10 % BOPs spending that time on extra LSQ+ instead beat refinement in every seed (+0.19 / +0.05 / +0.02 dB on average; smallest seed margin 0.001 dB). Our rule for practitioners is therefore: when fine-tuning follows, use the QUBO assignment and spend the compute on recovery, and keep refinement for deployment without fine-tuning. On the three public checkpoints, where you note that the search takes hours, refinement now adds only +0.03 to +0.07 dB, so at the search costs reported in the submission we would omit it there.

W2 / (b) — Is {6,7,8} too strict, and does it hold beyond the denoiser? We compared {4..8}, {5..8} and {6,7,8} on HalfUNet, the three public checkpoints and a ResNet-20 classifier on CIFAR-10, with the criteria fixed before running. At the reported 4.10 % budget the restriction costs nothing on any of the five models. A larger set helps mainly at tighter budgets: {5..8} is better on two public models around 2.9 % (+0.21 and +0.35 dB) and on ResNet-20 at 2.34 %, while on ResNet-20 the sets are within 0.25 points at every budget ≥ 2.93 %. On HalfUNet the objective does not need the restriction: an exact solver returns the same assignment for all three sets at every tested budget ≥ 2.93 %, and a same-γ check attributes the loss of the larger set under the submitted estimator to annealing error. So {6,7,8} is not a universal boundary. The revision will describe it as a search-space restriction that is free at the reported budget and should be re-checked for a new domain or a tighter budget.

W2 — Structural terms. We agree that both are hand-crafted priors, not graph-derived, and the revision will call them that. Their measured effect is small: isolated with an exact solver, skip compatibility changes validation PSNR by −0.05 to +0.09 dB, and the module-order term is inactive (it changes one assignment, by 0.0004 dB).

Reviewer_M7Y9OpenReview ↗

Feedback to authors rebuttal

Thanks for authors great effort for rebuttal. My concerns have been addressed; I shall keep my score.

Thank you for taking the time to read our rebuttal. We’re glad it addressed your concerns.

Reviewer_2nLzOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a task-aware QUBO framework for mixed-precision quantization, jointly optimizing weight and activation bit-width assignments under computational constraints. The proposed formulation incorporates gradient-based weight sensitivity, empirical activation sensitivity, bit-operation (BOP) costs, and selected structural interaction priors. To compensate for inaccuracies in the QUBO surrogate, the authors further introduce QUBO-seeded PROTES refinement, which directly evaluates candidate quantization configurations using task-specific validation metrics. Experiments on SIDD image denoising models demonstrate improvements in static quantization quality over uniform quantization and a HAWQ-style baseline.

Critical strengths

(1) This paper integrates separate weight and activation sensitivity estimates, computational costs, and structural interactions into a unified QUBO formulation.

(2) The QUBO-seeded PROTES refinement addresses potential discrepancies between estimated quantization damage and actual task performance.

Critical weaknesses

(1) The proposed framework primarily combines established components, including sensitivity-based quantization, QUBO optimization, and PROTES refinement. While their integration into joint weight–activation allocation is meaningful, the individual components are not fundamentally new.

(2) The main experiments primarily compare against uniform quantization and a HAWQ-style sensitivity–knapsack allocator. However, the baseline is an adaptation rather than the complete HAWQ-V3 method, and comparisons with differentiable mixed-precision search approaches, such as EdMIPS, are missing. Furthermore, exact or stronger discrete allocation baselines are not evaluated.

Questions for authors

(1) What are the essential conceptual differences from existing QUBO-based neural network compression methods that incorporate task-aware sensitivities or black-box refinement? Which proposed components contribute most significantly to the improvements?

(2) Given the marginal gains observed after recovery in the repeated-search experiment and the performance reversal under the tightest budget, could the authors investigate recovery-aware refinement objectives? Alternatively, could they compare their approach against allocating the same computational budget to additional recovery steps?

3 comments

Thank you for the constructive review. We ran the exact, controlled, multi-seed and compute-matched experiments you asked for. Numbers after fine-tuning are three-seed LSQ+ means with paired seeds, at strict BOP budgets.

W1 / Q1 — What is new, and which components matter. We agree that the components are established. The difference from earlier work is the problem they are combined to solve. Earlier QUBO work optimizes rounding at a fixed precision (Muñiz Subiñas et al.), combines pruning and quantization (Wang et al.) or prunes channels (Orabi et al.), and CLADO, OMPQ and EMQ allocate weight bits only, under a model-size budget. We allocate weight and activation precision per layer, jointly, under a BOP budget.

As for the components, two matter before fine-tuning. PROTES refinement adds +0.19 to +0.71 dB test PSNR (see Q2). In a controlled ablation on HalfUNet (exact solver, same candidate set, both objectives with a validation-tuned W/A balance), our damage terms lead the Hessian-trace term by +0.13 to +0.33 dB test PSNR. The skip constraint changes validation PSNR by only −0.05 to +0.09 dB, and the module-order term is inactive. After LSQ+, our allocation (annealed, native balance) leads the tuned Hessian-trace allocator by only 0.00–0.08 dB, which is why we claim the formulation and solver, not better sensitivity terms.

W2 — Exact and stronger baselines.

  • HAWQ-V3. As published it ties each layer's weight and activation precision (INT4 or INT8), so it does not search our joint W/A space. On HalfUNet with our quantizers and skip constraint, tied INT4/INT8 is 4.9 to 6.3 dB below uniform in test PSNR before fine-tuning (details in our reply to Reviewer y5GF). We therefore relabel the submitted row "Hessian-trace knapsack" and add a version with an exact solve and a validation-tuned W/A balance.
  • Exact solutions. The surrogate solves exactly as a MILP in under 2 s, even for 260 layers. For the submitted surrogate, annealing was within 0.82 % of the proven optimum at the tight budgets. For the corrected surrogate, the exact optimum differs from our annealed row by −0.03 to +0.18 dB validation PSNR before fine-tuning, and its fine-tuning is scheduled.
  • Differentiable search. We ported EdMIPS's search to our candidate set and quantizers.

PSNR after LSQ+ (dB):

BOP budget2.93 %3.42 %4.10 %4.69 %
Uniform36.5536.9537.0037.19
Submitted HAWQ-style baseline36.6136.8637.1037.19
Hessian-trace, tuned balance36.9737.0137.0937.15
FIT-based, tuned balance, our solver37.0137.1037.1437.21
EdMIPS (ported)36.92*37.1137.1637.23
Ours (annealed)36.9837.0837.1437.15

*at 2.78 % BOPs, its only search under that budget.

The calibrated allocators end within about 0.1 dB of each other. Ours beats uniform and the submitted baseline in every seed at 2.93–4.10 % (+0.13 to +0.43 dB over uniform), but at 4.69 % both are 0.04 dB ahead of ours. Two caveats apply to the rows: ours are annealed and leave 0.5–1.1 % of each budget unused, and EdMIPS updates network weights and took ten GPU searches of 11–13 minutes each (our cap; none met our convergence criterion).

Q2 — Refinement after fine-tuning. Your concern is confirmed: PROTES adds +0.19 to +0.71 dB test PSNR before fine-tuning but changes it by only −0.07 to +0.06 dB after LSQ+. Of your two suggestions we took the second and did not test a recovery-aware refinement objective. At 2.93 / 3.42 / 4.10 % we spent the PROTES search time (about 490–500 s, 47 extra epochs, learning rate chosen on validation) on more LSQ+ instead, which beat refinement in every seed, by +0.19 / +0.05 / +0.02 dB on average (smallest seed margin 0.001 dB). The revision will therefore say that, at the tested budgets, compute is better spent on recovery when fine-tuning follows, and that refinement is for deployment without fine-tuning. The 2.93 % reversal in Table 1 was within recovery noise, so the table will report three-seed means ± SD.

Correction. While answering W2 we found that the weight-error estimator overstated 7–8-bit error. The revision uses the corrected estimator and a strict BOP budget for every method, and keeps the submitted results in the appendix.

Thanks for the rebuttal and additional experiments. The authors have addressed my concerns regarding methodological clarification and component contributions (W1), and PROTES refinement after recovery (Q2).

The concerns regarding the baseline comparisons (W2) remain insufficiently addressed. First, EdMIPS fails to meet the convergence criterion, limiting the reliability of the comparison. Second, BOP budgets are not strictly matched, where the proposed annealed solutions leave 0.5–1.1% of the requested budgets unused, compromising the fairness of the comparison. Third, MILP establishes optimality only for the QUBO surrogate, not actual network performance, while fine-tuning results for the corrected exact solutions remain pending.

AuthorsOpenReview ↗

Follow-up on baseline comparisons

Thank you for the follow-up and for confirming that W1 and Q2 are addressed.

EdMIPS convergence. We agree that the ported comparison has limited reliability. Expected BOP cost, task loss, and selector entropy stabilized, but none of the ten searches met our final assignment stability criterion (Appendix C.3). We therefore treat this result as provisional. After LSQ+ fine-tuning, the port scores 0.02–0.04 dB above Ours (exact) at the three budgets from 3.42% to 4.69% (Table 2).

Budget matching. The 0.5–1.1% unused budget in our first reply refers to Ours (annealed). The revised main comparison uses Ours (exact), tuned FIT, and tuned Hessian-trace under the same hard BOP limits. Each uses at least 99.95% of every primary-split budget (Appendix D, Table 12). After fine-tuning, tuned FIT is within 0.04 dB of Ours (exact) and tuned Hessian-trace within 0.08 dB across those budgets (Table 2). The EdMIPS port still underspends, so the same budget-matching claim does not apply to that row.

What the MILP certifies. The certificate applies to the damage surrogate within solver tolerance. We assess network quality separately. Fine-tuning of the corrected exact assignments is complete, and Table 2 reports three paired LSQ+ seeds. Against the best fitting uniform configuration, Ours (exact) gains 0.17 and 0.13 dB at 2.93% and 3.42%, respectively, in all three seeds. The mean gains are below 0.01 dB at 4.10% and 4.69%. We have narrowed the paper’s claims accordingly.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of the revised PDF

We now recommend exact hard-budget allocation followed by fine-tuning, supported by a controlled evaluation of when joint weight/activation allocation helps. The revision retains the submission's problem, models, data, candidate sets and form of the damage terms, with changes listed in Section 1's last paragraph and Appendix F, Table 23.

What still stands. Against the best fitting uniform on HalfUNet’s primary split, Ours (exact) gains 0.17 and 0.13 dB at 2.93 and 3.42%, and 0.53 dB at the off-grid 2.539%, in all three seeds after fine-tuning. The comparator is the eligible uniform configuration with the highest mean test PSNR after fine-tuning, a choice that favours uniform. Gains are below 0.01 dB at 4.10 and 4.69% on the primary split, while the source-disjoint gain is 0.57 dB at 4.10% (Sections 5.1 and 5.3). On the primary split after fine-tuning, no tested allocator beats Ours (exact) by more than 0.04 dB, and profiling plus the exact solve takes about 7 s (Sections 5.1 and 5.4). We therefore claim the formulation and its evaluation, not a better sensitivity signal.

Where the requests are answered. Sections 1–3 clarify terms and equations. Section 2 now positions the formulation against integer programs over per-layer weight/activation pairs. Appendices A–C give notation and protocols. Section 5.1 adds exact and stronger baselines, Section 5.2 static tests including Restormer, Section 5.3 the source-disjoint split, Section 5.4 compute-matched extra fine-tuning and costs, and Section 5.5 ablations and ResNet-20.

Corrections and scope. We corrected a weight-error estimator that overstated 7–8-bit error and also fed the original Hessian-trace baseline. Appendix F explains the fix and keeps the original results (Table 24). Using the best fitting uniform lowers several gains we reported against the uniform configuration at the budget in our first replies. Refinement improves static quality, while its advantage after fine-tuning is inconsistent on the primary split and mixed on the source-disjoint split. It starts from the annealed assignment. Refinement from the exact assignment is untested. On public CNNs, FIT-based (tuned) matches refinement statically. BOPs cover allocated layers. Inference speedup was not measured. Code will be released with the final version.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.