The DOI matches its Crossref record; 71 of the 72 references it cites were found in a registry or an index; 11 of the 19 sources its claims cite were checked for support.
HADQ-Net: A Power-Efficient and Hardware-Adaptive Deep Convolutional Neural Network Translator Based on Quantization-Aware Training for Hardware Accelerators
Open accessAbstract
With the increasing demand for implementing deep-learning models on devices on resource-constrained devices, the development of power-efficient neural networks has become imperative. This paper introduces HADQ-Net, a novel framework for optimizing deep convolutional neural networks (CNNs) through Quantization-Aware Training (QAT). By compressing 32-bit floating-point (FP32) precision weights and activation values to lower bit-widths, HADQ-Net significantly reduces memory footprint and computational complexity while maintaining high accuracy. We propose adaptive quantization limits based on the statistical properties of each layer or channel, coupled with normalization techniques, to enhance quantization efficiency and accuracy. The framework includes algorithms for QAT, quantized convolution, and quantized inference, enabling efficient deployment of deep CNN models on edge devices. Extensive experiments across tasks such as super-resolution, classification, object detection, and semantic segmentation demonstrate the trade-offs between accuracy, model size, and computational efficiency under various quantization levels. Our results highlight the superiority of QAT over post-training quantization methods and underscore the impact of quantization types on model performance. HADQ-Net achieves significant reductions in memory footprint, computational complexity, and energy consumption, making it ideal for resource-constrained environments without sacrificing performance.
Overview
CoChat Summary
## TL;DR
HADQ-Net is a quantization-aware training framework for compressing deep CNNs to lower bit-widths while preserving accuracy and improving efficiency for hardware accelerators and edge devices. The paper reports that QAT outperforms PTQ across bit-widths, and that the method reduces parameter size from 41.94 MB at FP32 to 1.31 MB at 1-bit/1-bit.
## Key findings
- On CIFAR-100 with a pre-trained ResNet18, the proposed method achieved accuracy of 0.633, 0.751, 0.797, 0.862, 0.901, 0.907, 0.908, 0.910 at 1/1 through 8/8 bit-widths, compared with 0.496, 0.549, 0.562, 0.679, 0.849, 0.862, 0.875, 0.875 for PTQ.
- The paper states that QAT consistently outperforms PTQ across all bit-width configurations.
- Parameter size decreases from 41.94 MB at FP32 to 1.31 MB at 1/1, with intermediate sizes of 2.62, 3.93, 5.24, 6.55, 7.86, 9.17, and 10.49 MB for 2/2 through 8/8.
- In the fixed- vs mixed-precision comparison, mixed precision reduced weight size to 3.27 MB while achieving 0.823 accuracy, versus 3.93 MB and 0.797 accuracy for fixed precision.
- For classification, the paper reports results such as ResNet18: 0.594 to 0.960 across 1/1 to 8/8, and DenseNet121: 0.588 to 0.941 across the same range.
- For detection, faster_rcnn reached 0.868 to 0.972 mAP across low-to-full precision, and for segmentation AP@IoU 0.50:0.95 rose from 0.708 at 4/4 to 0.787 at FP32.
## Methods
- The study proposes HADQ-Net, a quantization-aware training (QAT) framework using uniform quantization, with dynamic activation quantization during inference and experiments on layer-wise vs channel-wise and fixed vs mixed precision quantization.
- The quantization algorithm uses adaptive clipping limits based on mean and standard deviation, followed by normalization and quantization; STE is used for backpropagation, and bias quantization is omitted.
- Experiments were run in PyTorch on an NVIDIA GeForce RTX 3090 Ti (24 GB), with multiple retrainings; datasets named in the text include CIFAR-100, ImageNet Object Localization, Hymenoptera, Berkeley Segmentation, and coco-2017.
## Limitations
- The paper states that QAT introduces significant training overhead, which remains a challenge for practical deployment.
- The authors note that their current results focus on training outcomes, and that comprehensive inference performance metrics will require implementation on dedicated edge computing platforms.
- The text indicates that some evaluations used a small subset of ImageNet, and the paper itself says this may not fully reflect real-world performance.
Model-written summary. Check it against the paper before relying on it.
Integrity signals
Registry
Cites
Statistics
Verification
Claims the paper makes
7 Valid · 1 Not valid · 3 Not checked · 79 Own results and interpretations
Every source the paper cites for this claim was checked and supports it.
In the paper
“However, traditional quantization methods often suffer from accuracy degradation, particularly when using low-precision representations.”
- Also stated in: § 2. Literature Review · plain
“However, quantization below 8-bit often leads to significant accuracy loss, requiring careful handling [17,22].”
What the sources say
- Supportedcites [17] Krishnamoorthi, R. (2018). Quantizing deep convolutional networks for efficient inference: A whitepaper. arXiv. source
Section 3.2.4 (p. 22) reports substantially better accuracy with per-channel than per-layer quantization at 4 bits, while the difference at 8 bits was not significant. It also reports accuracy losses from lower-precision activation quantization. These findings support the claim that low-precision quantization can degrade accuracy, though the paper describes techniques that mitigate the loss. (LLM, conf=0.90)
- Supportedcites [22] Choukroun, Y., Kravchik, E., Yang, F., and Kisilev, P. (2019, January 27–28). Low-bit Quantization of Neural Networks for Efficient Inference. Proceedings of the IEEE/CVF International Conference on Computer Vision Workshop (ICCVW), Seoul, Republic of Korea. source
The Related Work section states that quantizing full-precision weights and activations to fewer than 8 bits usually causes significant accuracy loss. This directly supports the claim about accuracy degradation at low precision. (LLM, conf=0.94)
plainstates a cause- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Also stated in: § 2. Literature Review · plain
The sources with a verdict support this claim; the rest weren’t checked yet, have no full text we could read, or their verdicts aren’t rated yet.
In the paper
“While PTQ is simpler and more efficient, QAT generally yields higher accuracy, particularly for lower bit-widths, when sufficient training data are available [14,15].”
What the sources say
- Not checked — no full text of the sourcecites [14] Xu, 2023, Improving Extreme Low-Bit Quantization With Soft Threshold, IEEE Trans. Circuits Syst. Video Technol. source
- Supportedcites [15] Wu, H., Judd, P., Zhang, X., Isaev, M., and Micikevicius, P. (2020). Integer Quantization for Deep Learning Inference: Principles and Empirical Evaluation. arXiv. source
Section 5.2 (p. 10) reports that QAT fine-tuning improved accuracy over PTQ in most evaluated cases. Related Work (p. 2) reports that training with quantization was required for high accuracy at lower bit-widths. The paper's QAT results come from fine-tuning on training datasets (Appendix A.2, p. 18), consistent with the claim's data-availability condition. (LLM, conf=0.86)
plainstates a cause- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
Every source the paper cites for this claim was checked and supports it.
In the paper
“QAT is important to achieve higher accuracy during inference, as it models quantization errors during training to match quantization effects during the inference [8].”
What the sources say
- Supportedcites [8] Jacob, B., Kligys, S., Chen, B., Zhu, M., Tang, M., Howard, A., Adam, H., and Kalenichenko, D. (2018, January 18–22). Quantization and training of neural networks for efficient integer-arithmetic-only inference. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, Salt Lake City, UT, USA. source
Section 3 (p. 5) says training simulates quantization effects in the forward pass to match quantized inference, while the Introduction (p. 2) says the framework minimizes accuracy loss from quantization. The Discussion (p. 9) says this training helps restore model accuracy. (LLM, conf=0.94)
plainstates a cause- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Supported
A result the paper reports from its own data — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“Extensive experiments across tasks such as superresolution, classification, object detection, and semantic segmentation demonstrate the trade-offs between accuracy, model size, and computational efficiency under various quantization levels.”
boostedstates an association- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
A result the paper reports from its own data — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“Our results highlight the superiority of QAT over post-training quantization methods and underscore the impact of quantization types on model performance.”
boostedstates a cause- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
33 more claims
Create a free account to see all 90 claims and to verify the 2 references still open.
- It may state two separate claims as one. 37
- Probably not a substantial claim. 25
- The quote only partly states this claim. 3
Citations
Cites
XNOR-Net: ImageNet Classification Using Binary Convolutional Neural Networks
Rastegari, M., Ordonez, V., Redmon, J., & Farhadi, A. (2016) · Lecture Notes in Computer Science · doi:10.1007/978-3-319-46493-0_32 · Cited by 4,408
Efficient Processing of Deep Neural Networks: A Tutorial and Survey
Sze, V., Chen, Y.-H., Yang, T.-J., & Emer, J. S. (2017) · Proceedings of the IEEE · doi:10.1109/jproc.2017.2761740 · Cited by 4,215
Quantization and Training of Neural Networks for Efficient Integer-Arithmetic-Only Inference
Jacob, B., Kligys, S., Chen, B., Zhu, M., Tang, M., Howard, A., Adam, H., & Kalenichenko, D. (2018) · 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition · doi:10.1109/cvpr.2018.00286 · Cited by 4,114
BinaryConnect: Training Deep Neural Networks with binary weights during propagations
Courbariaux, M., Bengio, Y., & David, J.-P. (2015) · arXiv · doi:10.48550/arxiv.1511.00363 · Cited by 2,180
EIE: efficient inference engine on compressed deep neural network
Han, S., Liu, X., Mao, H., Pu, J., Pedram, A., Horowitz, M. A., & Dally, W. J. (2016) · ACM SIGARCH Computer Architecture News · doi:10.1145/3007787.3001163 · Cited by 1,937
Estimating or Propagating Gradients Through Stochastic Neurons for Conditional Computation
Bengio, Y., Léonard, N., & Courville, A. (2013) · arXiv · doi:10.48550/arxiv.1308.3432 · Cited by 1,910
DoReFa-Net: Training Low Bitwidth Convolutional Neural Networks with Low Bitwidth Gradients
Zhou, S., Wu, Y., Ni, Z., Zhou, X., Wen, H., & Zou, Y. (2016) · arXiv · doi:10.48550/arxiv.1606.06160 · Cited by 1,690
Hubara, I., Soudry, D., & Yaniv, R. E. (2016) · arXiv · doi:10.48550/arxiv.1602.02505 · Cited by 1,667
A Brief Overview of ChatGPT: The History, Status Quo and Potential Future Development
Wu, T., He, S., Liu, J., Sun, S., Liu, K., Han, Q.-L., & Tang, Y. (2023) · IEEE/CAA Journal of Automatica Sinica · doi:10.1109/jas.2023.123618 · Cited by 1,452
A Survey of Quantization Methods for Efficient Neural Network Inference
Gholami, A., Kim, S., Dong, Z., Yao, Z., Mahoney, M. W., & Keutzer, K. (2022) · Low-Power Computer Vision · doi:10.1201/9781003162810-13 · Cited by 1,219
Research works printed but not linked: 3
- Choi, J., Venkataramani, S., Srinivasan, V., Gopalakrishnan, K., Wang, Z., and Chuang, P. (April, January 31). Accurate and Efficient 2-bit Quantized Neural Networks. Proceedings of the 2nd SysML Conference, Palo Alto, CA, USA.
- Dong, Z., Yao, Z., Gholami, A., Mahoney, M.W., and Keutzer, K. (November, January 27). HAWQ: Hessian aware quantization of neural networks with mixed-precision. Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), Seoul, Republic of Korea.
- Shen, X., Kong, Z., Yang, C., Han, Z., Lu, L., Dong, P., Lyu, C., Li, C., Guo, X., and Shu, Z. (2024). EdgeQAT: Entropy and distribution guided quantization-aware training for the acceleration of lightweight LLMs on the edge. arXiv.
Cited by
A Study on Image Recognition Technologies for Lightweight Autonomous Robots
Zhou, Y. (2026) · International Conference on Electronic Information Technology and Computer Engineering (EITCE) · doi:10.1109/eitce70137.2026.11634300
“To maximise system throughput on computationally constrained FPGA platforms, this study adopts the classical DoReFa-Net quantisation algorithm and implements an aggressive dual-quantisation strategy of 4-bit weights and 4-bit activations [11].”
Jariri, N., Allabouche, K., Benaly, M., Chaman, M., Majdoubi, R., & Hadjoudja, A. (2026) · International Conference on Innovative Research in Applied Science, Engineering and Technology (IRASET) · doi:10.1109/iraset68627.2026.11538436
“More recent studies explore mixed-precision strategies, extreme low-bit representations, and hardware-aware quantization techniques that explicitly consider architectural constraints [19], [20], [21], [22].”
A Hardware-Aware Analysis of PTQ and QAT Quantized CNNs for Object Detection on FPGA
Jariri, N., Allabouche, K., Benaly, M., Chaman, M., Majdoubi, R., & Hadjoudja, A. (2026) · Engineering Technology & Applied Science Research · doi:10.48084/etasr.17475 · Cited by 2
“To mitigate these effects, QAT has been proposed as a more robust alternative to PTQ [17, 18].”
Ye, X., Liu, Y., Zhao, L., et al. (2025) · Results in Engineering · doi:10.1016/j.rineng.2025.108905
Metadata
- Authors
- Oflamaz, C. U., & Yalçın, M. E.
- Affiliations
- Istanbul Technical University (TR)
- Department of Electronics and Communication Engineering, İstanbul Technical University, 34469 İstanbul, Türkiye
- Corresponding author
- Oflamaz, Can Uğur
- ORCID
- Oflamaz, Can Uğur
- Venue
- Electronics
- Publisher
- MDPI AG
- Published
- 2025-09-18
- Volume
- 14
- Issue
- 18
- Pages
- 3686
- ISSN
- 2079-9292 (ISSN-L)
- OpenAlex
- W4414334266
- Semantic Scholar
- 6cb3af9e180e70cdcb47ec005fd6cd3d4d3f6241
- Cited by
- 4
- References
- 47
- Citations by year
- 2026: 3 · 2025: 1
- FWCI
- 1.41
- Type
- journal article
- Language
- en
- Open access
- gold
- doi.org
- Licence
- Locations
- Electronics (publishedVersion, cc-by)
- Istanbul Technical University Academic Open Archive (Istanbul Technical University) (submittedVersion)
- Funders
- Aselsan Inc.
- Topics
- Physical Sciences › Computer Science › Computer Vision and Pattern Recognition › Advanced Neural Network Applications
- Physical Sciences › Engineering › Electrical and Electronic Engineering › Low-power high-performance VLSI design
- Physical Sciences › Computer Science › Artificial Intelligence › Domain Adaptation and Few-Shot Learning
- Keywords
- quantization-aware training, convolutional neural networks, neural network compression, edge devices, power efficiency, hardware accelerators, memory footprint reduction, energy consumption
About these checks
Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.
The claim verdicts this report counts are experimental: their accuracy is not yet measured.
A language model drew the claims from the paper on 2026-10-10, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).
A second model labelled each claim on 2026-10-10 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).
Cited sources are checked with a language model’s help (gpt-6-sol).
Statistics and arithmetic are recomputed, not judged.
Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.
What counts as a claim ›How we measure confidence ›How we check this ›
Are you an author, or do you see a problem with this page? Request a review
Data sources
Bibliographic data: Semantic Scholar, OpenAlex, Crossref.
Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).
Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.