Talk
Talk by Mohd Aquib
Title: Disentangled and Interpretable Representation Learning for Effective Facial Expression Recognition Systems: Methodological Overview and Future Perspectives Teams link
Abstract:
Facial expression recognition is challenging because expression-related information is often entangled with identity, pose, illumination, and other subject-specific variations. Moreover, conventional deep learning models typically provide limited insight into their predictions and may produce overconfident decisions for subtle, ambiguous, or overlapping expressions. This talk presents my research on developing effective, interpretable, and robust facial expression recognition systems through disentangled representation learning and uncertainty-aware AI.
The talk will first discuss adaptive variational autoencoder-based frameworks for separating expression-relevant representations from identity and other nuisance factors. It will then present an information-theoretic approach for learning identity-independent expression representations without requiring explicit identity annotations. Subsequently, I will describe a semi-supervised guided adaptive β-VAE integrated with an interval Type-2 fuzzy logic classifier. In this framework, disentangled expression representations are translated into interpretable fuzzy rules, while interval-valued memberships explicitly capture uncertainty associated with ambiguous facial expressions. I will also discuss recent methods for adaptively balancing reconstruction quality and disentanglement in variational autoencoders through game-theoretic and reinforcement-learning-based strategies.
The talk will conclude with future research directions based on Ashby Neural Networks for affective computing. This proposed framework combines differentiable Tversky similarity, General Recognition Theory, and stochastic evidence accumulation to model facial expressions as noisy multidimensional percepts with explicit prototypes, perceptual overlap, decision regions, and uncertainty. The broader objective is to develop psychologically grounded, interpretable, and trustworthy affective computing frameworks for real-world image and video analysis.
Speaker Bio:Dr. Mohd Aquib received the Ph.D. degree in Electrical Engineering from the Indian Institute of Technology Kanpur in 2026. Prior to this, he received the M.Tech. degree in Control and Instrumentation Engineering and the B.Tech. degree in Electrical Engineering from Jamia Millia Islamia, New Delhi, in 2018 and 2015, respectively. He is a recipient of the prestigious Fellowship for Academic and Research Excellence (FARE) at IIT Kanpur and is currently serving as a FARE Fellow in the Department of Electrical Engineering. His research lies at the intersection of affective computing, disentangled representation learning, variational autoencoders, fuzzy systems, and interpretable and uncertainty-aware artificial intelligence. His research works has been published in reputed journals and conferences, including IEEE Transactions on Affective Computing, IEEE Transactions on Artificial Intelligence, IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops etc.
- 13 - 13 July,2026
- 11:00 am - 12:00 pm
- Online