Quantização Guiada pela Lei de Benford: Compressão Log-Uniforme de Pesos para Modelos de Visão Médica Eficientes

• Arthur Negrão UFOP
• Guilherme Silva UFOP
• Matheus Vieira UFOP
• Ederson N. F. G. Júnior UFOP
• Eduardo José da Silva Luz UFOP
• Pedro Silva UFOP

DOI: https://doi.org/10.5753/sbcas.2026.21342

Resumo

Benford-Quant, um método de quantização pós-treinamento guiado pela Lei de Benford, é avaliado na classificação de imagens médicas. O estudo considera ResNet-18, ResNet-50 e ViT-Base nos conjuntos BloodMNIST, PathMNIST e OrganCMNIST, com comparação contra RTN e NF4. Os resultados mostram que o BenQ preserva desempenho competitivo em F1-Score, com redução de memória de até 7,7x, evidenciando que seus benefícios não se restringem a LLMs, mesmo em cenários desfavoráveis como ruído gaussiano e variações de contraste, onde métodos não uniformes tendem a maior robustez. Esses achados posicionam o BenQ como uma alternativa viável para visão médica em ambientes com recursos limitados.

Referências

Abid, A., Sinha, P., Harpale, A., Gichoya, J., and Purkayastha, S. (2021). Optimizing medical image classification models for edge devices. In International Symposium on Distributed Computing and Artificial Intelligence, pages 77–87. Springer.

Askari Hemmat, M. H., Hemmat, R. A., Hoffman, A., Lazarevich, I., Saboori, E., Mastropietro, O., Sah, S., Savaria, Y., and David, J.-P. (2022). Qreg: On regularization effects of quantization. arXiv preprint arXiv:2206.12372.

Benford, F. (1938). The law of anomalous numbers. Proceedings of the American philosophical society, pages 551–572.

Dettmers, T., Pagnoni, A., Holtzman, A., and Zettlemoyer, L. (2023). Qlora: Efficient finetuning of quantized llms. Advances in neural information processing systems, 36:10088–10115.

Frantar, E., Ashkboos, S., Hoefler, T., and Alistarh, D. (2022). Gptq: Accurate post-training quantization for generative pre-trained transformers. arXiv preprint arXiv:2210.17323.

He, K., Zhang, X., Ren, S., and Sun, J. (2015). Deep residual learning for image recognition. CoRR, abs/1512.03385.

Hill, T. P. (1995). The significant-digit phenomenon. The American Mathematical Monthly, 102(4):322–327.

Kusk, M. W. and Lysdahlgaard, S. (2023). The effect of gaussian noise on pneumonia detection on chest radiographs, using convolutional neural networks. Radiography, 29(1):38–43.

Lu, Z.-x., Qian, P., Bi, D., Ye, Z.-w., He, X., Zhao, Y.-h., Su, L., Li, S.-l., and Zhu, Z.-l. (2021). Application of ai and iot in clinical medicine: summary and challenges. Current medical science, 41(6):1134–1150.

Manduva, V. C. (2024). Advancing ai in edge computing with graph neural networks for predictive analytics. The Metascience, 2(2):75–102.

Marchisio, K., Dash, S., Chen, H., Aumiller, D., Üstün, A., Hooker, S., and Ruder, S. (2024). How does quantization affect multilingual llms? In Findings of the Association for Computational Linguistics: EMNLP 2024, pages 15928–15947.

Mekala, A., Atmakuru, A., Song, Y., Karpinska, M., and Iyyer, M. (2025). Does quantization affect models’ performance on long-context tasks? In Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing, pages 9433–9481.

Nagel, M., Fournarakis, M., Amjad, R. A., Bondarenko, Y., Van Baalen, M., and Blankevoort, T. (2021). A white paper on neural network quantization. arXiv preprint arXiv:2106.08295.

Negrão, A., Silva, P., Freitas, V. L., Moreira, G., and Luz, E. (2026). Benford’s law as a distributional prior for post-training quantization of large language models. arXiv preprint arXiv:2602.00165.

Ott, J., Sun, H., Rinaldi, E., Mauro, G., Servadei, L., and Wille, R. (2025). Exploiting benford’s law for weight regularization of deep neural networks. Transactions on Machine Learning Research.

Sahu, S. K., Java, A., and Shaikh, A. (2021). On the connection of benford’s law and neural networks. CoRR.

Shandhi, M. M. H. and Dunn, J. P. (2022). Ai in medicine: Where are we now and where are we going? Cell Reports Medicine, 3(12).

Tong, Y., Yuan, J., and Hu, C. (2026). Enhancing quantization-aware training on edge devices via relative entropy coreset selection and cascaded layer correction. IEEE Transactions on Mobile Computing.

Wu, B., Xu, C., Dai, X., Wan, A., Zhang, P., Yan, Z., Tomizuka, M., Gonzalez, J., Keutzer, K., and Vajda, P. (2020). Visual transformers: Token-based image representation and processing for computer vision.

Xi, L., Li, C., Anari, M. S., and Rezaee, K. (2025). Integrating wearable health devices with ai and edge computing for personalized rehabilitation. Journal of Cloud Computing, 14(1):64.

Xu, X., Lu, Q., Yang, L., Hu, S., Chen, D., Hu, Y., and Shi, Y. (2018). Quantization of fully convolutional networks for accurate biomedical image segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8300–8308.

Yang, J., Shi, R., Wei, D., Liu, Z., Zhao, L., Ke, B., Pfister, H., and Ni, B. (2023). Medmnist v2-a large-scale lightweight benchmark for 2d and 3d biomedical image classification. Scientific data, 10(1):41.

Zhai, S., Wang, H., Sun, L., Zhang, B., Huo, F., Qiu, S., Wu, X., Ma, J., Wu, Y., and Duan, J. (2022). Artificial intelligence (ai) versus expert: A comparison of left ventricular outflow tract velocity time integral (lvot-vti) assessment between icu doctors and an ai tool. Journal of applied clinical medical physics, 23(8):e13724.

Zhao, H. (2023). Applications of embedded systems in medicine: Challenges and future trends. Highlights Sci. Eng. Technol, 62:31–35.