Faculty of Electronic Engineering, Niš, Serbia

Fakultet tehničkih nauka, Novi Sad, Sebija

Akademija tehničko vaspitačkih strukovnih studija Niš

Faculty of Electronic Engineering, Niš, Serbia

University of Manchester, United Kingdom

Faculty of Electronic Engineering, Niš, Serbia

The paper is available at the following link: PDF

Abstract:

Low-precision arithmetic has emerged as a promising strategy for enabling efficient deployment of deep learning models on resource-constrained platforms. While quantization has been extensively explored in vision and speech models, the effect of low-bit numerical formats on Graph Neural Networks (GNNs)—particularly for affective computing tasks remains insufficiently studied. Emotion recognition from conversational data requires capturing subtle semantic and relational cues, making precision reduction potentially challenging. This work investigates the impact of 8-bit floating- point (FP8) and 8-bit fixed-point (FxP8) quantization on the performance of GNN-based emotion recognition systems. Experiments are conducted on the IEMOCAP benchmark using the CogMEN architecture, a GNN model designed for conversational emotion understanding. Quantization is applied to model weights and activations during inference, and model performance is assessed using classification accuracy and macro- F1 score, both critical metrics for emotion recognition. The results show that FP8 quantization consistently outperforms FxP8, demonstrating superior accuracy retention due to its wider dynamic range and improved numerical stability. Although FxP8 exhibits a modest performance degradation, it remains a viable solution for highly resource constrained deployments owing to its simpler arithmetic and lower hardware complexity. These findings highlight the trade-offs between numerical robustness and implementation efficiency when deploying GNN-based emotion recognition systems on edge and embedded platforms.

Keywords:

Emotion recognition, Graph Neural Networks, FP8, FxP8, quantization, CogMEN, edge AI.

REFERENCES:

[1] Burkitt, Ian. "SOCIAL RELATIONSHIPS AND EMOTIONS." Sociology 31, no. 1 (1997): 37–55. http://www.jstor.org/stable/42855768.
[2] N. H. Frijda, "The Emotions", Cambridge University Press, 1986.
[3] C. Yang, X. Zhang, Y. Chen, Y. Li, S. Yu, B. Zhao, T. Wang, L. Luo, S. Gao "Emotion dependet language featuring depression", J. Behav. Therapy Exp. Psych, 81,101883, 2023. DOI: 10.1016/j.jbtep.2023.101883
[4] S. Mahlke, M. Minge, "Emotions and EMG measures of facial muscules in interactive contexts", Cong. Emot. 6, 169-200, 2006
[5] N. Antoniou, A. Katsamanis, T. Giannakopoulos and S. Narayanan, "Designing and Evaluating Speech Emotion Recognition Systems: A Reality Check Case Study with IEMOCAP," ICASSP 2023 - 2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), Rhodes Island, Greece, 2023, pp. 1-5, doi: 10.1109/ICASSP49357.2023.10096808.
[6] Z. Perić, B. Denić, M. Dinčić, S. Perić, "An Approximate Closed-Form Expression for Calculating Performance of Floating-Point Format for the Laplacian Source", Informatica, Vol. 36, No. 1, pp. 125-140, 2025. https://doi.org/10.15388/25-INFOR587
[7] D. Lin, S. Talathi, S. Annapureddy, "Fixed point quantization of deep neural networks", Proceedings of 33rd International Conference on Machine Learning, DOI: https://doi.org/10.48550/arXiv.1511.06393
[8] R. Goyal, J. Vanschoren, V. van Acht, S. Nijssen, "Fixed-point Quantization of Convolutional Neural Networks for Quantized Inference on Embedded Platforms". Mathematics and Computer ScienceData Mining, al. arXiv.org. https://arxiv.org/abs/2102.02147
[9] N. Mitschke, M. Geizmann, K. Noffz, R. Wittmann, "A Fixed-Point Quantization Technique for Convolutional Neural Networks Based on Weight Scaling", 2019 IEEE International Conference on Image Processing (ICIP), DOI:10.1109/ICIP.2019.8803490
[10] B. Denić, Z. Perić, M. Dinčić, S. Perić, N. Simić, M. Anđelković, "Switched 32-Bit Fixed-Point Format for Laplacian-Distributed Data", Information 16, no.7: 574, July 2025. https://doi.org/ 10.3390/info16070574
[11] N. Simic, S. Suzic, T. Nosek, M. Vujovic, Z. Peric, M. Savic, V. Delic, "Speaker recognition using constrained convolutional neural networks in emotional speech", Entropy, 24(3), 2022,DOI:10.3390/e24030414
[12] N. Simic, S. Suzic, N. Milosevic, V. Stanojev, T. Nosek, B. Popovic, D. Bajovic "Enchancing emotion recognition through federated learning: a multimodal approach with convolutional neural networks", Applied Sciences.14(4), 1325, 2024, DOI:10.3390/app14041325
[13] A. Joshi, A. Bhat, A. Jain, A. V. Modi, "COGMEN: Contextualized GNN based multimodal emotion recognition",In: Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Seatle, USA, 4148-4164, 2022
[14] T. Ðurkić, N. Simić, S. Suzić, D. Bajović, Z. Perić, V. Delić, "Multimodal Emotion Recognition Using Compressed Graph Neural Networks", Speech and Computer. SPECOM 2024. Lecture Notes in Computer Science, vol 15300. Springer, Cham. https://doi.org/10.1007/978-3-031-78014-1_9
[15] N. Majumder, S. Poria, D. Hazarika, R. Mihalcea, A. Gelbukh, E. Cambria, "DialogueRNN: An Attentive RNN for Emotion Detection in Conversations", July 2019Proceedings of the AAAI Conference on Artificial Intelligence 33(01):6818-6825, DOI: 10.1609/aaai.v33i01.33016818