Task-Aware QUBO Allocation for Mixed-Precision Quantization
Osama Orabi, Artur Zagitov, Hadi Salloum, Viktor A. Lobachev, Yaroslav Kholodov
cs.LG · cs.CV
Quick verdict
QUBO+PROTES mixed-precision allocation yields small gains (≤0.1 dB post-recovery, single runs) on a denoiser; candid reporting, modest contribution.
Summary
The paper casts mixed-precision bit-width allocation for weights and activations as a QUBO surrogate combining sensitivity-weighted quantization damage, an analytical bit-operation (BOP) cost, and hand-selected structural priors, solved with simulated annealing; the chosen route seeds PROTES, a tensor-train probabilistic optimizer that refines the allocation by direct validation evaluation, followed by LSQ+ recovery. On a custom NAFBlock-based HalfUNet denoiser trained on SIDD patches, the refined route reaches 37.192 dB PSNR after LSQ+ at 4.035% routed-layer BOPs, versus 37.092 dB at 4.101% for a HAWQ-style baseline. A repeated-search experiment finds QUBO-seeded PROTES improves on QUBO alone by only 0.046 dB statically and 0.007 dB after recovery, so LSQ+ largely closes the refinement gap, while NAFNetSmall retains a larger gain (33.703 to 34.079 dB). The authors report routing instability, negative ablations, and test-set hyperparameter sweeps with unusual candor. Key limitations: headline comparisons rest on single recovery runs, with gaps (up to about 0.1 dB) smaller than the 0.208 dB recovery variation observed on one fixed route; the primary patch-random SIDD split lets source images cross train/test partitions; deployment is fake-quantized floating-point execution in which quantized inference is slower than FP32 (6.735 versus 3.002 ms on an H100), so BOP savings remain analytical. No code release is mentioned.
Review complete
Assessment at a glance
5
risk flags identified
● high 0
● medium 3
● low 2
4 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- A task-aware QUBO surrogate with separate weight/activation sensitivity profiles, BOP cost and structural priors produces joint W+A allocations competitive with a HAWQ-style sensitivity–knapsack baseline under a shared quantization/recovery stack.
- The refined route reaches 37.192 dB after LSQ+ at 4.035% routed-layer BOPs versus 37.092 dB at 4.101% for the HAWQ-style baseline on the HalfUNet denoiser.
- In the primary repeated-search experiment, LSQ+ recovery closes nearly all of the gap between QUBO allocation and expensive direct refinement (0.046 dB static, 0.007 dB recovered), making refinement's value architecture-dependent (NAFNetSmall retains 0.376 dB).
- Recovery changes route ordering and the return on allocation search, so recovery and optimization expense belong in the allocation decision.
- Deployment measurements characterize a fake-quantized FP implementation with no latency or size benefit; BOP reductions are analytical allocation savings only.
Strengths
Unusually candid negative-result and stability reporting
Section 5.6 and Appendices D.9–D.10 document that solver seeds change routes materially (coordinate agreement 0.28–0.31, PSNR SD >1 dB at two γ values, Table 15) and that validation-subset changes yield only 0.37–0.58 joint routing agreement (Table 14); D.18 lists five failed alternatives. Table 9 explicitly maps each experiment to its evidence scope, preventing protocol pooling.
Honest deployment accounting
Section 5.7/Table 13 state plainly that the reference implementation is fake-quantized FP: quantized GPU inference is 6.735 ms vs 3.002 ms FP32 on H100, serialized size (≈322 KB) and peak allocation do not improve, and the abstract concedes BOP reductions are analytical only. No hardware speedup is claimed.
Shared-stack baseline design
Reproducibility-oriented detail
Weaknesses
Headline effects sit below the protocol's own noise floor
Leaky, tiny primary test bed
No realized efficiency
Baseline and ablation scope limits
Test-set hyperparameter selection
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
The paper proposes a two-stage mixed-precision quantization (MPQ) pipeline. Stage 1 is a task-aware QUBO surrogate that allocates weight bits from {4..8} and activation bits from {6,7,8} per routed convolution, combining candidate reconstruction damage scaled by sensitivity scores (Fisher-inspired EMA gradient energy for weights; downstream 4-bit probe response for activations), a BOP compute term (γ), skip-connection compatibility penalties and module-order interaction priors (ω=0.15), solved by simulated annealing with one-hot repair. Stage 2 seeds PROTES (tensor-train probabilistic search) from the QUBO route and refines it by direct validation evaluation under an excess-compute penalty (Eq. 4), followed by LSQ+ recovery. Central claims: the refined route reaches 37.192 dB at 4.035% routed-layer BOPs vs 37.092 dB at 4.101% for a HAWQ-style baseline (Table 1); LSQ+ largely closes the QUBO-vs-refinement gap (0.007 dB, Table 3); refinement's value is architecture-dependent (NAFNetSmall +0.376 dB); and deployment is fake-quant with analytical-only BOP savings. Table 9 maps every study to its protocol — an unusually explicit evidence-scope contract.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 4 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What does this paper propose for mixed-precision quantization?
It assigns each convolution a weight bit-width (4–8) and activation bit-width (6–8) by solving a QUBO — a binary quadratic optimization that trades sensitivity-weighted quantization damage against an analytical bit-operation (BOP) compute cost, plus structural priors. The QUBO route, found by simulated annealing, then seeds PROTES, a tensor-train probabilistic optimizer that fine-tunes the allocation by directly evaluating candidate networks on validation data, before LSQ+ quantization-aware recovery training.
Does the quantized model actually run faster or smaller?
No, not in this paper's implementation. Measurements use fake quantization with floating-point execution: on an H100, quantized inference took 6.735 ms versus 3.002 ms for FP32, and serialized model size and GPU memory did not improve. The reported ~96% BOP reductions are analytical savings that would require native low-bit kernels and packed weight storage to realize.
How much better is the QUBO approach than simpler alternatives?
In the main experiment, the refined QUBO+PROTES route scored 37.192 dB PSNR at 4.035% BOPs versus 37.092 dB at 4.101% for a HAWQ-style baseline and 37.132 dB for QUBO alone — gaps of about 0.06–0.1 dB from single recovery runs. A repeated-search experiment found refinement added only 0.007 dB over QUBO alone after recovery, though a second architecture (NAFNetSmall) retained a larger 0.376 dB gain.
Is code available for this paper?
The paper does not mention a code release. It does, however, document seeds, hyperparameters, QUBO matrix construction, per-stage timings and environment details (PyTorch 2.9.1, CUDA 12.8, H100) across its appendices, which would aid reimplementation.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.