A Systematic Analysis of Multi-Modal Neural Network Architectures for Emotion Recognition

دسته بندی موضوعی : برنامه ریزی درسی و آموزشی

زبان : انگلیسی

نوع فایل : PDF

تعداد صفحات : 8

وضعیت : این مقاله توسط هیات داوران پایگاه ایران ژورنال داوری شده است
فصل نامه علمی تخصصی پژوهش برتر - سال اول شماره سوم - پاییز 1404

نویسندگان :
Raheleh rezazadeh 1PhD Student in Artificial Intelligence (Computer Science), Islamic Azad University, Sabzevar ra.rezazadeh@iau.ir

ثبت شده توسط : ایران ژورنال

لینک مستقیم مقاله : https://iran-journals.com/doc/1231

چکیده فارسی :

Intelligent learner emotion recognition in modern educational environments is identified as a key driver for enhancing engagement and personalizing content. As interactions within virtual platforms become increasingly complex, traditional unimodal models lose efficacy due to high sensitivity to environmental noise and uncertainty in single-source data. This study aims to provide a systematic analysis of advanced multi-modal architectures during 2024–2025, integrating heterogeneous data—including visual, auditory, and textual inputs—to reconstruct learners’ psychological states with high fidelity.
In the methodology layer, the research focuses on novel neural structures leveraging tri-attention mechanisms in hyper-attentive transformers, shared Conformer layers for handling incomplete data, and deep reinforcement learning algorithms. The analysis demonstrates that mid-level data fusion at the tensor level, combined with large language models such as BERT alongside vision transformers, produces more resilient models capable of extracting and interpreting nonlinear correlations across modalities, even under conditions of facial occlusion or audio signal fluctuations.
Findings confirm that the selected architectures, achieving accuracy rates exceeding 93% and inference latencies below 25 milliseconds, provide an efficient infrastructure for real-time educational interventions. Evaluations on benchmark datasets such as AffectNet and ECLS-K highlight the high potential of these systems to improve learning processes for students with special needs and optimize instructional strategies in scalable cloud-based environments. Ultimately, this study proposes a framework for integrating emotion-aware artificial intelligence into large-scale educational systems, promising transformative impacts on educational equity and the quality of digital learning.

کلمات کلیدی فارسی :
Multimodal Emotion Recognition, Deep Reinforcement Learning (DRL), Vision Transformers, Personalized Learning, Heterogeneous Data Fusion.