MUMBAI, India, Sept. 22 -- Intellectual Property India has published a patent application (202644101871 A) filed by C Pavani; and Priya Gupta on August 21, 2026, for Metacognitive Gated Multimodal Fusion System For Explainable Affective Computing.

Inventor includes Priya Gupta.

The application for the patent was published on September 18, 2026, under issue no. 38/2026.

Abstract: The present invention relates to an Explainable Metacognitive Gated Multimodal Fusion System for Affective Computing, designed for robust and transparent multimodal emotion recognition in noisy and uncertain real-world environments. The system addresses limitations of conventional multimodal fusion approaches, particularly their inability to dynamically assess the reliability of individual modalities and their lack of transparent decision-making. The proposed system processes textual, acoustic, and visual inputs using specialized deep learning modules comprising a RoBERTa-based text encoder, WavLM-based acoustic encoder, and R(2+1)D- 18 visual encoder, and maps the extracted representations into a standardized latent feature space. A metacognitive gating network evaluates modality-specific confidence information and dynamically assigns trust weights to the respective modalities before fusion, thereby suppressing unreliable or noisy inputs and emphasizing reliable affective cues. The fused representation is subsequently processed to generate an emotion prediction. An integrated Explainable AI module comprising gating activation maps and SHAP-based analysis provides interpretable evidence regarding modality contribution and feature importance. Experimental validation on the MELD dataset demonstrated an accuracy of approximately 64% and a weighted F1-score of 63.00%, while recovering the detection of the fear emotion from a unimodal F1-score of 0.00 to 0.21 through multimodal complementary cues. The system achieved a mean five-fold validation accuracy of 79.83% with a standard deviation of 0.94 and an average inference latency of 0.6228 ms per sample, demonstrating robustness, explainability, and computational efficiency. The proposed invention is suitable for applications including adaptive education, human-computer interaction, psychological monitoring, healthcare support, and real-time affective computing.

Disclaimer: Curated by HT Syndication.