Pith. sign in

REVIEW 4 major objections 4 minor 155 references

A layer's quantization sensitivity in an LLM depends on upstream bitwidths; MixQuant marginalizes errors over random quantized contexts, yielding budget-agnostic scores that beat every baseline tested.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-01 03:45 UTC pith:UGGK55AP

load-bearing objection Genuinely new idea about context-dependent layer sensitivity, empirically strong but with fixable evidence gaps—worth a serious referee. the 4 major comments →

arxiv 2607.23047 v1 pith:UGGK55AP submitted 2026-07-25 cs.LG cs.AI

MixQuant: Adaptive Mixed-Precision Quantization for Large Language Models

Ashitabh Misra , Madhav Agrawal , Arham Jain , Tarek Abdelzaher This is my paper
classification cs.LG cs.AI
keywords mixed-precision quantizationadaptive quantizationpost-training quantizationlarge language modelsmean-field approximationbit allocationgreedy allocationcontext-aware sensitivity
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

This paper tries to show that mixed-precision quantization of large language models is being scored in the wrong context. Existing adaptive methods measure each layer's sensitivity against a full-precision (FP16) network, but at deployment every layer is quantized, and the paper demonstrates that a layer's NMSE and downstream allocation shift by up to two to three orders of magnitude depending on the bitwidths of its upstream layers. MixQuant's remedy is a mean-field marginalization: each layer's distortion is averaged over random quantized upstream configurations, giving budget-agnostic scores, plus quantizer parameters calibrated on plans the allocator itself produces and a tail penalty that steers spare budget away from the lowest bitwidths. If these claims hold, a single offline calibration can serve any memory budget at deployment with a cheap greedy solve, improving average downstream accuracy by up to 8 points and cutting perplexity from 12.43 to 10.70 at the tightest budget across Llama-3.2-3B, Llama-2-7B, and Mistral-7B under both AWQ and GPTQ.

Core claim

The central discovery is that per-layer sensitivity scores computed on the FP16 model are systematically unrepresentative of the fully quantized networks actually deployed: the output distortion of a module, and even the bit allocation it induces through a solver, varies by up to two to three orders of magnitude across different quantized upstream contexts. MixQuant replaces the unknown deployment context with a random prior: it draws upstream bitwidths independently and uniformly from the admissible bit set, measures each module's normalized mean-squared error under K random contexts, and averages. The resulting decoupled distortion table depends only on (module, bitwidth), so it is budget-

What carries the argument

The load-bearing object is the decoupled distortion score, the expected normalized mean-squared error of quantizing module ℓ to b bits when upstream bitwidths are drawn i.i.d. from the uniform prior over the admissible bit set, estimated by K=10 Monte Carlo draws. This mean-field marginalization makes each module's score independent of the other modules' assigned bitwidths, which in turn makes the ensuing multiple-choice knapsack objective additive and budget-agnostic. Supporting machinery: plan-aware technique parameters, where the feasible budget range is bucketed and each bucket's parameters are averaged over anchor plans the greedy solver produces; and the tail regularizer, which discoun

Load-bearing premise

The load-bearing premise is that averaging each module's distortion over ten independently, uniformly drawn upstream bitwidths correctly ranks module upgrades under the actual deployment distribution, which is neither independent nor uniform — it is produced by the greedy solver under a memory budget, with many modules pinned at the floor bitwidth, and the paper gives no analysis of this mismatch.

What would settle it

Re-run the MixQuant pipeline but sample the upstream contexts in Stage 1 from plans the greedy allocator itself produces across the budget range, instead of from the uniform prior; if the resulting bit allocations and downstream accuracies shift materially, or if increasing K from 10 to several hundred changes the distortion table, the mean-field marginalization is not genuinely budget-agnostic. A second check: test at a budget far outside the range used to build the Stage-2 parameter buckets and see whether the advantage over FP16-scored baselines collapses, which would indicate the plan-awar

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • One offline calibration serves any memory budget at deployment: a new budget triggers a single O(L|B| log L) greedy pass over the precomputed distortion table, with no re-calibration.
  • The largest gains appear where quantization is most aggressive — at the tightest budgets many modules sit at the floor and upstream drift is largest — so the method is most valuable for low-memory edge deployment.
  • The framework wraps any base quantizer whose parameters can be recalibrated on a given context; the paper demonstrates this for AWQ and GPTQ, so other PTQ techniques should inherit the same benefit.
  • Scoring against the FP16 network is a correctable bias: even strong FP16-scored metrics paired with an exact ILP solver are outperformed by the context-averaged scores, and the greedy allocation matches the ILP's downstream quality at a fraction of the cost.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The uniform prior over upstream bitwidths is likely a loose surrogate for the contexts the greedy solver actually produces — at a given budget many modules are pinned at the floor, not spread uniformly — so an allocator-aware prior (sampling contexts from solver-generated plans) could rank upgrades more accurately at extreme budgets; this is a directly testable variant.
  • The context-dependence finding suggests a broader design principle for LLM quantization: any module whose input statistics are perturbed by upstream compression (activation quantization, KV-cache compression, low-rank approximations) should be scored under those perturbed contexts rather than against a full-precision reference; testing this on activation-only quantization would be a natural next e
  • The tail penalty saturates beyond γ≈5, and the mechanism implies the penalty changes results only when a few modules would otherwise monopolize spare budget; comparing plan bitwidth distributions at γ=0 and γ=10 would let practitioners set γ by inspection rather than by sweep.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

4 major / 4 minor

Summary. The paper proposes MixQuant, an adaptive mixed-precision quantization framework for LLMs. Stage 1 computes per-module distortion scores by averaging NMSE over K random quantized upstream contexts drawn from a uniform prior. Stage 2 calibrates base-quantizer parameters (GPTQ Hessians, AWQ scales/clipping) on anchor plans generated by the same greedy solver used at deployment. Stage 3 adds a tail penalty that discourages leaving modules at the lowest bitwidths. A single greedy multiple-choice knapsack solve then serves any memory budget at deployment. The evaluation covers Llama-3.2-3B, Llama-2-7B, and Mistral-7B with AWQ and GPTQ across three budgets per model, reporting consistent improvements over FP16-scored metrics, LIM, and CoopQ, plus a component ablation and a tail-penalty sensitivity sweep.

Significance. If the claims hold, MixQuant is a practically valuable contribution: one offline calibration supports arbitrary deployment budgets, and the reported gains at tight budgets are large (e.g., 7–8 average accuracy points on Llama-3.2-3B at 1.25 GB). The paper has real strengths: the component ablation (Table 5) shows that the distortion table carries signal and that the greedy solver matches an ILP on the surrogate objective; the tail-penalty sweep (Figure 4) shows a saturating benefit; and the evaluation is broad in models, base quantizers, and budgets. However, the central theoretical device—the uniform-prior mean-field score—is not validated against the actual distribution of contexts produced by the allocator, and the empirical claims lack any error bars or multiple-seed analysis. These issues need to be addressed before the contribution can be fully credited as stated.

major comments (4)
  1. [§4.1, Eq. (4)] The uniform prior π(b_{<ℓ}) = ∏ Unif(B) is an unvalidated assumption. Figure 2 documents that per-module NMSE varies by 2–3 orders of magnitude across contexts, and Figure 2b shows that the induced allocations vary with context. Yet the paper never compares the distortion table D_π with one computed under the empirical distribution of contexts generated by Algorithm 1 at the actual deployment budgets. The component ablation (Table 5) demonstrates that the score table carries signal relative to no scores or reversed scores, but it does not test whether the uniform weighting is the right weighting. Please add (i) a direct comparison of allocations and downstream metrics under the uniform prior vs. an allocator-induced context distribution, and (ii) a sensitivity analysis over the prior (e.g., different marginal bitwidth distributions or an empirical prior). Without this, the claim that mea
  2. [§5, Tables 2–4] All results are single point estimates with no error bars, multiple seeds, or confidence intervals. This matters because the scoring procedure is stochastic: K=10 Monte Carlo draws per (module, bitwidth) pair (§4.1), and the MC estimator is unlikely to resolve the 2–3 orders-of-magnitude variation shown in Figure 2a. Since the paper repeatedly claims improvement 'in every setting,' please report means and variances over independent MC seeds and, if feasible, multiple calibration runs. A single run cannot support claims built on margins as small as 0.8 points (Table 3, Llama-2-7B at 4.0 GB under GPTQ).
  3. [§5, baselines paragraph] LIM and CoopQ are evaluated on bitwidth set B={3,...,8}, while the paper itself acknowledges they 'were designed and validated on small bitwidth sets, typically {2,3,4}.' Figure 3 confirms they collapse to the extremes of B. This makes the headline claim 'outperforms adaptive and mixed-precision baselines in every setting' less informative: for these two baselines the comparison is outside their validated operating range. Please either evaluate LIM and CoopQ on their native bitwidth sets (and clearly report the corresponding memory budgets), or explicitly restrict the claim to methods designed for wide bitwidth sets. The controlled FP16-prior baselines are informative, but the LIM/CoopQ comparison needs to be reframed or re-run.
  4. [§4.2 and Table 5] The plan-aware parameter calibration is stated as a contribution, but its individual effect is never isolated. The text acknowledges that for AWQ the averaging of parameter sets 'is a heuristic' (§4.2), and Table 5 ablates only the full Stage 1+Stage 2 pipeline versus score-free and reversed allocations. There is no comparison of plan-aware µ versus FP16-calibrated µ, nor a sensitivity study over the number of intervals I and anchor plans J. Given that this component is claimed to exploit the fact that 'the technique and the solver are fixed before deployment,' please add an ablation that turns Stage 2 on/off or compares against FP16 calibration, and report sensitivity to I and J.
minor comments (4)
  1. [Figure 4] The y-axis labels (0, -0.1, -0.3, -1, -3) appear to be on a reciprocal or log scale but are not labeled as such. Please clarify the axis scale or use a standard log axis.
  2. [Table 5] The row label 'Mem. budget' is terse and could be misread as the memory budget itself. Consider renaming to 'Budget-only (no scores)' for clarity.
  3. [§3 and Algorithm 1] Notation is overloaded: b denotes a full plan in §3, a single module bitwidth in Eq. (5)–(6), and the current plan in Algorithm 1, while b^+ denotes a successor bitwidth. Please make the level of the variable explicit throughout.
  4. [§5] The text alternates between 'WikiText-2' and 'Wikitext' (e.g., 'Wikitext PPL' in Table 1). Please standardize the spelling.

Circularity Check

0 steps flagged

No circularity found: core claims are measured on held-out tasks and the only self-citation is non-load-bearing.

full rationale

MixQuant's derivation chain is self-contained: Stage 1 computes d_bar_l(b) by Monte Carlo averaging over a stated uniform prior (Eq. 4), Stage 2 calibrates technique parameters on anchor plans produced by the same greedy solver, and Stage 3 adds a tuned tail penalty, with final allocations evaluated on held-out WikiText-2 validation and EleutherAI tasks. No fitted constant is renamed as a prediction: the reported accuracy and perplexity gains are measured outcomes, not algebraic consequences of the calibration quantities. The only self-citation (Misra et al. 2025 for the MCKP formulation, Section 3) appears alongside two independent citations (Yao et al. 2020, Li et al. 2023) and is not load-bearing; the formulation is standard and the paper's contribution does not rest on it. The uniform-prior mean-field choice and the possible mismatch with the solver-induced context distribution are robustness and correctness concerns, not circularity, because the paper does not define its predicted quantities in terms of that prior and the evaluation is external to the calibration objective. The Stage 2 calibration-on-solver-plans loop is a heuristic self-consistency procedure rather than a reduction of the reported gains to the calibration inputs.

Axiom & Free-Parameter Ledger

4 free parameters · 4 axioms · 0 invented entities

The ledger shows that the method's main free choices are the uniform prior, the Monte Carlo sample size, the tail penalty, and the plan-bucketing hyperparameters. These are set by hand and not derived from first principles; the empirical gains depend on them, though the sensitivity sweep for γ mitigates that particular choice. The uniform prior is the most consequential ad hoc assumption because it defines the entire distortion table.

free parameters (4)
  • Tail penalty γ = 10
    Chosen by hand (Section 5, 'unless stated otherwise ... tail penalty γ=10'). The sensitivity sweep (Figure 4) shows the effect saturates around γ≈5, but no selection criterion (e.g., validation-based tuning) is reported. The value 10 sits in the flat region, but the choice is not derived.
  • Monte Carlo draws K = 10
    Chosen by hand (Section 5, 'K=10 Monte Carlo context draws per (module, bitwidth) pair'). No sensitivity analysis is reported. Figure 2a shows NMSE/KL variation of 2–3 orders of magnitude across contexts, implying high variance in the marginal estimate; K=10 may be too small for stable scores.
  • Budget intervals I and anchor plans J = I=3, J=5
    Chosen by hand (Section 4.2, 'partition the feasible memory range into I intervals', 'sample J budgets ... yields J anchor plans'). No sensitivity analysis is reported. The bucketing heuristic determines how plan-aware the technique parameters are.
  • Uniform prior π over upstream bitwidths = Unif(B) for each upstream module
    The marginalization distribution in Eq. (4) assumes independent uniform bitwidths for every upstream module. This is a modelling assumption, not derived from deployment conditions or the solver's output distribution. The choice of uniform over B is arbitrary and affects all distortion scores.
axioms (4)
  • domain assumption The deployment loss is well-approximated by the sum of per-module mean-field decoupled NMSE scores (Section 3, Eq. 2).
    The true network loss is replaced with an additive MCKP surrogate. The paper states the additive objective presumes independence; averaging over a prior does not guarantee that the sum of marginal scores is a faithful proxy for the actual network loss under a specific plan.
  • ad hoc to paper Upstream bitwidths are independent and uniformly distributed over B (Section 4.1, Eq. 4).
    Deployed plans are memory-budget-constrained and produced by the greedy solver, so upstream bitwidths are neither independent nor uniform. The uniform prior is chosen for tractability and is not justified by the deployment distribution.
  • domain assumption Simulated quantization (quantize–dequantize to FP16) is a faithful proxy for real integer hardware inference (Section 5, 'Quantization is simulated').
    All reported memory and accuracy numbers are based on simulation; no actual hardware deployment or integer kernel validation is performed. Real inference may introduce additional errors (e.g., non-linear operations, mixed-precision kernels) not captured here.
  • domain assumption The base quantizer's technique parameters can be recalibrated under arbitrary quantized upstream contexts (Section 4.1, 'the technique parameters (µ) of the quantized modules are recalibrated under that context').
    For GPTQ this is Hessian recomputation on drifted activations, which is well-defined. For AWQ, it requires re-estimating smoothing factors and clipping ranges on quantized inputs; the paper does not specify how this is done or whether it is stable across contexts.

pith-pipeline@v1.3.0-alltime-deepseek · 18674 in / 13409 out tokens · 129850 ms · 2026-08-01T03:45:47.759918+00:00 · methodology

0 comments
read the original abstract

Mixed-precision quantization improves the accuracy of post-training quantization by allocating higher bitwidths to sensitive layers, but existing methods solve the allocation for a single fixed memory budget. In practice the budget varies across deployments and is unknown at calibration time. Adaptive quantization addresses this with one offline calibration that serves any budget, yet current methods score layer sensitivity in a manner that does not consider its dependency on quantization levels of other layers. We show that a layer's sensitivity depends strongly on the bitwidths of its upstream layers and that this dependence shifts the resulting preferred bit allocation. We propose MixQuant, a technique-agnostic adaptive framework that wraps any base quantizer. MixQuant marginalizes each layer's distortion over random quantized upstream configurations to obtain budget-agnostic scores, calibrates the quantizer's parameters on plans the allocator itself produces, and penalizes allocations that leave layers at the lowest bitwidths. A single greedy pass then serves any budget at deployment. Across Llama-3.2-3B, Llama-2-7B, and Mistral-7B under AWQ and GPTQ, MixQuant outperforms adaptive and mixed-precision baselines in every setting, improving average accuracy by up to 8 points and reducing perplexity from 12.43 to 10.70 at the tightest budget, while matching an ILP solver at negligible deployment cost.

Figures

Figures reproduced from arXiv: 2607.23047 by Arham Jain, Ashitabh Misra, Madhav Agrawal, Tarek Abdelzaher.

Figure 1
Figure 1. Figure 1: Overview of MIXQUANT. Stage 1 builds a budget-agnostic distortion table D ∈ R L×|B| by scoring each (module, bitwidth) pair under K random quantized upstream contexts (Section 4.1). Stage 2 estimates the technique parameters µ on activations induced by anchor plans the greedy solver produces across the feasible budget range, rather than on FP16 activations (Section 4.2). Stage 3 adds a tail penalty γ that … view at source ↗
Figure 2
Figure 2. Figure 2: Effect of upstream context, measured over the same 20 random quantized contexts in both [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Per-linear-layer bitwidth allocations by method. For Llama-3.2-3B (1.25 GB), Llama-2- [PITH_FULL_IMAGE:figures/full_fig_p010_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Effect of the tail penalty. Relative change in WikiText-2 perplexity versus the unregular [PITH_FULL_IMAGE:figures/full_fig_p011_4.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

155 extracted references · 21 canonical work pages · 3 internal anchors

  1. [1]

    Wood , title =

    Brooks Paige and Frank D. Wood , title =. Proceedings of the 31th International Conference on Machine Learning,. 2014 , url =

    2014

  2. [2]

    2023 , url =

    Li, Shiyao and Ning, Xuefei and Hong, Ke and Liu, Tengxuan and Wang, Luning and Li, Xiuhong and Zhong, Kai and Dai, Guohao and Yang, Huazhong and Wang, Yu , booktitle =. 2023 , url =

    2023

  3. [3]

    Approximate Computing: A Survey , year=

    Xu, Qiang and Mytkowicz, Todd and Kim, Nam Sung , journal=. Approximate Computing: A Survey , year=

  4. [4]

    PLDI , year=

    Reactive probabilistic programming , author=. PLDI , year=

  5. [5]

    Density-Based Semantics for Reactive Probabilistic Programming

    Guillaume Baudart and Louis Mandel and Christine Tasson , title =. CoRR , volume =. 2023 , url =. doi:10.48550/ARXIV.2308.01676 , eprinttype =. 2308.01676 , timestamp =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2308.01676 2023

  6. [6]

    Advances in Variational Inference , journal =

    Cheng Zhang and Judith B. Advances in Variational Inference , journal =. 2019 , url =. doi:10.1109/TPAMI.2018.2889774 , timestamp =

    arXiv 2019

  7. [7]

    2022 , url =

    Guillaume Baudart and Louis Mandel and Reyyan Tekin , editor =. 2022 , url =. doi:10.1145/3519941.3535066 , timestamp =

    arXiv 2022

  8. [8]

    2020 Design, Automation

    Hiromitsu Awano and Masanori Hashimoto , title =. 2020 Design, Automation. 2020 , url =. doi:10.23919/DATE48585.2020.9116302 , timestamp =

    arXiv 2020

  9. [9]

    Rutenbar and Tsuhan Chen , title =

    Claire Fang Fang and Rob A. Rutenbar and Tsuhan Chen , title =. 2003 International Conference on Computer-Aided Design,. 2003 , url =. doi:10.1109/ICCAD.2003.1257675 , timestamp =

    Pith/arXiv arXiv 2003

  10. [10]

    Rozier and Johann Schumann , editor =

    Johannes Geist and Kristin Y. Rozier and Johann Schumann , editor =. Runtime Observer Pairs and Bayesian Network Reasoners On-board FPGAs: Flight-Certifiable System Health Management for Embedded Systems , booktitle =. 2014 , url =. doi:10.1007/978-3-319-11164-3\_18 , timestamp =

    work page doi:10.1007/978-3-319-11164-3 2014

  11. [11]

    Eric Atkinson and Charles Yuan and Guillaume Baudart and Louis Mandel and Michael Carbin , title =. Proc. 2022 , url =. doi:10.1145/3563347 , timestamp =

    work page doi:10.1145/3563347 2022

  12. [12]

    Bayesian Robot Programming , journal =

    Olivier Lebeltel and Pierre Bessi. Bayesian Robot Programming , journal =. 2004 , url =. doi:10.1023/B:AURO.0000008671.38949.43 , timestamp =

    arXiv 2004

  13. [13]

    2013 , url=

    Fixed-Point Arithmetic: An Introduction , author=. 2013 , url=

    2013

  14. [14]

    2017 , url =

    Rajesh Ranganath , title =. 2017 , url =

    2017

  15. [15]

    Banerjee and Zbigniew T

    Subho S. Banerjee and Zbigniew T. Kalbarczyk and Ravishankar K. Iyer , editor =. AcMC \(. Proceedings of the Twenty-Fourth International Conference on Architectural Support for Programming Languages and Operating Systems,. 2019 , url =. doi:10.1145/3297858.3304019 , timestamp =

    arXiv 2019

  16. [16]

    Flexpoint: An Adaptive Numerical Format for Efficient Training of Deep Neural Networks , booktitle =

    Urs K. Flexpoint: An Adaptive Numerical Format for Efficient Training of Deep Neural Networks , booktitle =. 2017 , url =

    2017

  17. [17]

    2022 , booktitle =

    Chandra et al, Kartik , title =. 2022 , booktitle =

    2022

  18. [18]

    Journal of the American statistical Association , year=

    Variational inference: A review for statisticians , author=. Journal of the American statistical Association , year=

  19. [19]

    Zixin Huang and Saikat Dutta and Sasa Misailovic , title =. Innov. Syst. Softw. Eng. , volume =. 2022 , url =. doi:10.1007/S11334-021-00433-3 , timestamp =

    work page doi:10.1007/s11334-021-00433-3 2022

  20. [20]

    fpm library , author=

  21. [21]

    Jacob Laurel and Rem Yang and Shubham Ugare and Robert Nagel and Gagandeep Singh and Sasa Misailovic , title =. Proc. 2022 , url =. doi:10.1145/3563324 , timestamp =

    work page doi:10.1145/3563324 2022

  22. [22]

    Jacob Laurel and Rem Yang and Gagandeep Singh and Sasa Misailovic , title =. Proc. 2022 , url =. doi:10.1145/3498718 , timestamp =

    work page doi:10.1145/3498718 2022

  23. [23]

    Design Automation Conference (DAC) , year=

    Statheros: Compiler for Efficient Low-Precision Probabilistic Programming , author=. Design Automation Conference (DAC) , year=

  24. [24]

    Proceedings of the Twenty-Third International Conference on Architectural Support for Programming Languages and Operating Systems,

    Ruizhe Cai and Ao Ren and Ning Liu and Caiwen Ding and Luhao Wang and Xuehai Qian and Massoud Pedram and Yanzhi Wang , editor =. Proceedings of the Twenty-Third International Conference on Architectural Support for Programming Languages and Operating Systems,. 2018 , url =. doi:10.1145/3173162.3173212 , timestamp =

    arXiv 2018

  25. [25]

    Goodman , title =

    Daniel Ritchie and Paul Horsfall and Noah D. Goodman , title =. CoRR , volume =. 2016 , url =. 1610.05735 , timestamp =

    Pith/arXiv arXiv 2016

  26. [26]

    Continualization of Probabilistic Programs With Correction , booktitle =

    Jacob Laurel and Sasa Misailovic , editor =. Continualization of Probabilistic Programs With Correction , booktitle =. 2020 , url =. doi:10.1007/978-3-030-44914-8\_14 , timestamp =

    work page doi:10.1007/978-3-030-44914-8 2020

  27. [27]

    and Henzinger, Thomas A

    Gordon, Andrew D. and Henzinger, Thomas A. and Nori, Aditya V. and Rajamani, Sriram K. , title =. Future of Software Engineering Proceedings , pages =. 2014 , isbn =. doi:10.1145/2593882.2593900 , abstract =

    arXiv 2014

  28. [28]

    Proceedings of the 55th Annual Design Automation Conference,

    Shubham Jain and Swagath Venkataramani and Vijayalakshmi Srinivasan and Jungwook Choi and Pierce Chuang and Leland Chang , title =. Proceedings of the 55th Annual Design Automation Conference,. 2018 , url =. doi:10.1145/3195970.3196012 , timestamp =

    arXiv 2018

  29. [29]

    arXiv: Learning , year=

    Training deep neural networks with low precision multiplications , author=. arXiv: Learning , year=

  30. [30]

    An Unbiased

    Shuanglong Liu and Grigorios Mingas and Christos. An Unbiased. 2017 , url =. doi:10.1109/TC.2016.2630682 , timestamp =

    arXiv 2017

  31. [31]

    CoRR , volume =

    Nimish Shah and Laura Isabel Galindez Olascoaga and Wannes Meert and Marian Verhelst , title =. CoRR , volume =. 2021 , url =. 2103.00216 , timestamp =

    Pith/arXiv arXiv 2021

  32. [32]

    Neurocomputing , volume =

    Nico Piatkowski and Sangkyun Lee and Katharina Morik , title =. Neurocomputing , volume =. 2016 , url =. doi:10.1016/J.NEUCOM.2015.01.091 , timestamp =

    work page doi:10.1016/j.neucom.2015.01.091 2016

  33. [33]

    Ko and Yuji Chai and Rob A

    Glenn G. Ko and Yuji Chai and Rob A. Rutenbar and David Brooks and Gu. Accelerating Bayesian Inference on Structured Graphs Using Parallel Gibbs Sampling , booktitle =. 2019 , url =. doi:10.1109/FPL.2019.00033 , timestamp =

    arXiv 2019

  34. [34]

    Lebeck , title =

    Xiangyu Zhang and Ramin Bashizade and Yicheng Wang and Cheng Lyu and Sayan Mukherjee and Alvin R. Lebeck , title =. CoRR , volume =. 2020 , url =. 2003.04223 , timestamp =

    Pith/arXiv arXiv 2020

  35. [35]

    and Vainio, O

    Saramaki, T. and Vainio, O. , booktitle=. Structures for generating polynomial responses , year=

  36. [36]

    Bertrand Jeannet and Antoine Min. Apron:. Computer Aided Verification, 21st International Conference,. 2009 , url =. doi:10.1007/978-3-642-02658-4\_52 , timestamp =

    doi:10.1007/978-3-642-02658-4 2009

  37. [37]

    Vechev , editor =

    Timon Gehr and Sasa Misailovic and Martin T. Vechev , editor =. Computer Aided Verification - 28th International Conference,. 2016 , url =. doi:10.1007/978-3-319-41528-4\_4 , timestamp =

    doi:10.1007/978-3-319-41528-4 2016

  38. [38]

    IEEE Standard for Floating-Point Arithmetic , year=

    , journal=. IEEE Standard for Floating-Point Arithmetic , year=

  39. [39]

    Compiling Stan to generative probabilistic languages and extension to deep probabilistic programming , booktitle =

    Guillaume Baudart and Javier Burroni and Martin Hirzel and Louis Mandel and Avraham Shinnar , editor =. Compiling Stan to generative probabilistic languages and extension to deep probabilistic programming , booktitle =. 2021 , url =. doi:10.1145/3453483.3454058 , timestamp =

    arXiv 2021

  40. [40]

    Journal of Statistical Software , author=

    Stan: A Probabilistic Programming Language , volume=. Journal of Statistical Software , author=. 2017 , pages=. doi:10.18637/jss.v076.i01 , abstract=

    doi:10.18637/jss.v076.i01 2017

  41. [41]

    Automatically improving accuracy for floating point expressions , booktitle =

    Pavel Panchekha and Alex Sanchez. Automatically improving accuracy for floating point expressions , booktitle =. 2015 , url =. doi:10.1145/2737924.2737959 , timestamp =

    arXiv 2015

  42. [42]

    CoRR , volume =

    Yu Cheng and Duo Wang and Pan Zhou and Tao Zhang , title =. CoRR , volume =. 2017 , url =. 1710.09282 , timestamp =

    Pith/arXiv arXiv 2017

  43. [43]

    and Jankowiak, Martin and Obermeyer, Fritz and Pradhan, Neeraj and Karaletsos, Theofanis and Singh, Rohit and Szerlip, Paul and Horsfall, Paul and Goodman, Noah D

    Bingham, Eli and Chen, Jonathan P. and Jankowiak, Martin and Obermeyer, Fritz and Pradhan, Neeraj and Karaletsos, Theofanis and Singh, Rohit and Szerlip, Paul and Horsfall, Paul and Goodman, Noah D. , title =. J. Mach. Learn. Res. , month = jan, pages =. 2019 , issue_date =

    2019

  44. [44]

    Deep Learning with Limited Numerical Precision , booktitle =

    Suyog Gupta and Ankur Agrawal and Kailash Gopalakrishnan and Pritish Narayanan , editor =. Deep Learning with Limited Numerical Precision , booktitle =. 2015 , url =

    2015

  45. [45]

    and Krishnan, S

    Ramalingam, A. and Krishnan, S. , title =. Trans. Info. For. Sec. , month =. 2006 , issue_date =. doi:10.1109/TIFS.2006.885036 , abstract =

    arXiv 2006

  46. [46]

    CoRR , volume =

    Jonathan Long and Evan Shelhamer and Trevor Darrell , title =. CoRR , volume =. 2014 , url =. 1411.4038 , timestamp =

    Pith/arXiv arXiv 2014

  47. [47]

    Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

    Bit-Mixer: Mixed-precision networks with runtime bit-width selection , author=. Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

  48. [48]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    Adabits: Neural network quantization with adaptive bit-widths , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  49. [49]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    Instance-aware dynamic neural network quantization , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

  50. [50]

    Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

    Eq-net: Elastic quantization neural networks , author=. Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

  51. [51]

    Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision , pages=

    Improved Techniques for Quantizing Deep Networks with Adaptive Bit-Widths , author=. Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision , pages=

  52. [52]

    Latency-Constrained Input-Aware Quantization of Time Series Inference Workflows at the Edge , year=

    Misra, Ashitabh and Saoda, Nurani and Abdelzaher, Tarek , booktitle=. Latency-Constrained Input-Aware Quantization of Time Series Inference Workflows at the Edge , year=

  53. [53]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    One-shot model for mixed-precision quantization , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  54. [54]

    International Conference on Machine Learning , pages=

    Sdq: Stochastic differentiable quantization with mixed precision , author=. International Conference on Machine Learning , pages=. 2022 , organization=

    2022

  55. [55]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    Haq: Hardware-aware automated quantization with mixed precision , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

  56. [56]

    NeurIPS ML for Systems workshop, 2018 , year=

    Releq: an automatic reinforcement learning approach for deep quantization of neural networks , author=. NeurIPS ML for Systems workshop, 2018 , year=

    2018

  57. [57]

    Le , title =

    Barret Zoph and Quoc V. Le , title =. 5th International Conference on Learning Representations,. 2017 , url =

    2017

  58. [58]

    Search What You Want: Barrier Panelty

    Haibao Yu and Qi Han and Jianbo Li and Jianping Shi and Guangliang Cheng and Bin Fan , editor =. Search What You Want: Barrier Panelty. Computer Vision -. 2020 , url =. doi:10.1007/978-3-030-58545-7\_1 , timestamp =

    work page doi:10.1007/978-3-030-58545-7 2020

  59. [59]

    2021 , url =

    Dilin Wang and Meng Li and Chengyue Gong and Vikas Chandra , title =. 2021 , url =. doi:10.1109/CVPR46437.2021.00635 , timestamp =

    arXiv 2021

  60. [60]

    CoRR , volume =

    Bichen Wu and Yanghan Wang and Peizhao Zhang and Yuandong Tian and Peter Vajda and Kurt Keutzer , title =. CoRR , volume =. 2018 , url =. 1812.00090 , timestamp =

    Pith/arXiv arXiv 2018

  61. [61]

    Temporal Dynamic Quantization for Diffusion Models , booktitle =

    Junhyuk So and Jungwon Lee and Daehyun Ahn and Hyungjun Kim and Eunhyeok Park , editor =. Temporal Dynamic Quantization for Diffusion Models , booktitle =. 2023 , url =

    2023

  62. [62]

    Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

    Insta-bnn: Binary neural network with instance-aware threshold , author=. Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

  63. [63]

    European Conference on Computer Vision , pages=

    Cadyq: Content-aware dynamic quantization for image super-resolution , author=. European Conference on Computer Vision , pages=. 2022 , organization=

    2022

  64. [64]

    Bashima Islam and Shahriar Nirjon , title =. Proc. 2020 , url =. doi:10.1145/3411808 , timestamp =

    work page doi:10.1145/3411808 2020

  65. [65]

    2022 , url =

    Kwanghee Choi and Martin Kersner and Jacob Morton and Buru Chang , title =. 2022 , url =. doi:10.1109/ICASSP43922.2022.9747908 , timestamp =

    arXiv 2022

  66. [66]

    Proceedings of the 8th ACM/IEEE Conference on Internet of Things Design and Implementation , pages=

    LightEQ: On-Device Earthquake Detection with Embedded Machine Learning , author=. Proceedings of the 8th ACM/IEEE Conference on Internet of Things Design and Implementation , pages=

  67. [67]

    Bronstein and Uri C

    Moran Shkolnik and Brian Chmiel and Ron Banner and Gil Shomron and Yury Nahshan and Alexander M. Bronstein and Uri C. Weiser , title =. CoRR , volume =. 2020 , url =. 2002.07686 , timestamp =

    Pith/arXiv arXiv 2020

  68. [68]

    MultiQuant: Training Once for Multi-bit Quantization of Neural Networks , url =

    Xu, Ke and Feng, Qiantai and Zhang, Xingyi and Wang, Dong , year =. MultiQuant: Training Once for Multi-bit Quantization of Neural Networks , url =. doi:10.24963/ijcai.2022/504 , booktitle =

    work page doi:10.24963/ijcai.2022/504 2022

  69. [69]

    Straightening Out the Straight-Through Estimator: Overcoming Optimization Challenges in Vector Quantized Networks , booktitle =

    Minyoung Huh and Brian Cheung and Pulkit Agrawal and Phillip Isola , editor =. Straightening Out the Straight-Through Estimator: Overcoming Optimization Challenges in Vector Quantized Networks , booktitle =. 2023 , url =

    2023

  70. [70]

    Estimating or Propagating Gradients Through Stochastic Neurons for Conditional Computation , journal =

    Yoshua Bengio and Nicholas L. Estimating or Propagating Gradients Through Stochastic Neurons for Conditional Computation , journal =. 2013 , url =. 1308.3432 , timestamp =

    Pith/arXiv arXiv 2013

  71. [71]

    Classification of Radar Targets using Features Based on Warped Discrete Fourier Transform , url =

    Bujakovic, Dimitrije and Andric, Milenko and Bondzulic, Boban and Simic, Slobodan , year =. Classification of Radar Targets using Features Based on Warped Discrete Fourier Transform , url =. doi:10.23919/eusipco55093.2022.9909909 , booktitle =

    arXiv 2022

  72. [72]

    , author=

    Wavelet Transform Based Mel-scaled Features for Acoustic Scene Classification. , author=. INTERSPEECH , volume=

  73. [73]

    D. G. Bhalke and C. B. Rama Rao and Dattatraya S. Bormane , title =. J. Intell. Inf. Syst. , volume =. 2016 , url =. doi:10.1007/S10844-015-0360-9 , timestamp =

    work page doi:10.1007/s10844-015-0360-9 2016

  74. [74]

    Progressive Neural Compression for Adaptive Image Offloading Under Timing Constraints , url=

    Wang, Ruiqi and Liu, Hanyang and Qiu, Jiaming and Xu, Moran and Guérin, Roch and Lu, Chenyang , year=. Progressive Neural Compression for Adaptive Image Offloading Under Timing Constraints , url=. doi:10.1109/rtss59052.2023.00020 , booktitle=

    arXiv 2023

  75. [75]

    Howard and Menglong Zhu and Bo Chen and Dmitry Kalenichenko and Weijun Wang and Tobias Weyand and Marco Andreetto and Hartwig Adam , title =

    Andrew G. Howard and Menglong Zhu and Bo Chen and Dmitry Kalenichenko and Weijun Wang and Tobias Weyand and Marco Andreetto and Hartwig Adam , title =. CoRR , volume =. 2017 , url =. 1704.04861 , timestamp =

    Pith/arXiv arXiv 2017

  76. [76]

    and Zhu, Menglong and Zhmoginov, Andrey and Chen, Liang-Chieh , title =

    Sandler, Mark and Howard, Andrew G. and Zhu, Menglong and Zhmoginov, Andrey and Chen, Liang-Chieh , title =. CoRR , volume =. 2018 , url =

    2018

  77. [77]

    and Adam, Hartwig , title =

    Howard, Andrew and Sandler, Mark and Chu, Grace and Chen, Liang-Chieh and Chen, Bo and Tan, Mingxing and Wang, Weijun and Zhu, Yukun and Pang, Ruoming and Vasudevan, Vijay and Le, Quoc V. and Adam, Hartwig , title =. Proceedings of the IEEE International Conference on Computer Vision (ICCV) , year =

  78. [78]

    PyTorch: An Imperative Style, High-Performance Deep Learning Library , journal =

    Adam Paszke and Sam Gross and Francisco Massa and Adam Lerer and James Bradbury and Gregory Chanan and Trevor Killeen and Zeming Lin and Natalia Gimelshein and Luca Antiga and Alban Desmaison and Andreas K. PyTorch: An Imperative Style, High-Performance Deep Learning Library , journal =. 2019 , url =. 1912.01703 , timestamp =

    Pith/arXiv arXiv 2019

  79. [79]

    2024 33rd International Conference on Computer Communications and Networks (ICCCN) , pages=

    Acies-OS: A Content-Centric Platform for Edge AI Twinning and Orchestration , author=. 2024 33rd International Conference on Computer Communications and Networks (ICCCN) , pages=. 2024 , organization=

    2024

  80. [80]

    Advances in Neural Information Processing Systems , year =

    FOCAL: Contrastive Learning for Multimodal Time-Series Sensing Signals in Factorized Orthogonal Latent Space , author =. Advances in Neural Information Processing Systems , year =

Showing first 80 references.