MinMAE calibration method for convolutional neural network quantization
PeerJ Computer Science · 23.03.2026 · 12 · e3724
DOI: 10.7717/peerj-cs.3724 ↗Kurz gefasst
Fachartikel in der Zeitschrift PeerJ Computer Science, online erschienen am 23.03.2026. Verfasst von Artem Vasilev und 7 weiteren Autor:innen. Die Arbeit wurde bisher 2-mal zitiert (Crossref) und nennt 28 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0). Gefördert von: High-Performance Computing facilities at HSE University.
Abstract (Originalsprache)
This article introduces MinMAE, a novel activation calibration method for Post-Training Quantization (PTQ) that significantly reduces accuracy loss in Convolutional Neural Networks (CNN). Motivated by the need for high-fidelity quantization without costly retraining, MinMAE directly minimizes the Mean Absolute Error (MAE) between original and dequantized activations, making it robust to outliers that degrade standard methods. We evaluated MinMAE against absmax and mean+3std on YOLOv8n and YOLOv8s models. For 8-bit quantization on YOLOv8n, MinMAE achieved 8.4% mean Average Precision (mAP) loss, whereas standard methods led to drops of up to 14.4%. On the larger YOLOv8s model, MinMAE also demonstrated superior performance, reducing loss from 7.2% to just 5.2%. While this approach requires a higher one-time calibration time, the substantial accuracy gains establish MinMAE as a highly effective and broadly applicable strategy for improving the stability and performance of quantized CNNs.
Förderung
- High-Performance Computing facilities at HSE University
Zitieren
APA
Vasilev, A., Kapitanov, A. I., Solovyev, R., Telpukhov, D., Stempkovskiy, A., Kustov, A., Prutyanov, V., & Romanov, A. (2026). MinMAE calibration method for convolutional neural network quantization. PeerJ Computer Science, 12, e3724. https://doi.org/10.7717/peerj-cs.3724
Vancouver
Vasilev A, Kapitanov AI, Solovyev R, Telpukhov D, Stempkovskiy A, Kustov A, et al. MinMAE calibration method for convolutional neural network quantization. PeerJ Computer Science. 2026;12:e3724. doi:10.7717/peerj-cs.3724
BibTeX
@article{Vasilev2026,
author = {Vasilev, Artem and Kapitanov, Andrey I. and Solovyev, Roman and Telpukhov, Dmitry and Stempkovskiy, Alexander and Kustov, Alexander and Prutyanov, Viktor and Romanov, Aleksandr},
title = {MinMAE calibration method for convolutional neural network quantization},
journal = {PeerJ Computer Science},
year = {2026},
volume = {12},
pages = {e3724},
doi = {10.7717/peerj-cs.3724}
}
EUVIMED