CHAI Advanced PhD Student Talk: Rajesh Titung
CHAI Seminar Series
Refreshments will be served.
DATE: Monday, April 21, 2025
SPEAKER: Rajesh Titung, Advanced Ph.D. Candidate in Computing and Information Sciences, RIT
TITLE: Personalization and Generalization for Modeling Heterogeneous Affective Behaviors
IN PERSON: Golisano Hall (070), Room CYB-1710/1720
ABSTRACT: Communication and interaction involve complex emotional expressions with subtle variations shaped by sociocultural and interpersonal factors, making their perception ambiguous and open to valid variation in interpretation. This variability leads to heterogeneity in feature and label spaces and drives concept shifts between/within individuals, posing challenges for machine learning systems. Consequently, traditional centralized modeling, focused on generalization, often exhibits poor and variable performance across individuals in these perceptual prediction tasks, with limited generalizability to previously unseen individuals. This situation underscores two critical insights: first, the overemphasis on generalization at the expense of personalization and the need for principles to guide decisions on when to personalize or generalize; and second, the necessity of personalized models that effectively capture individualized behaviors. To address this research gap, we introduce an agreement-based taxonomy for strategic modeling decisions and a novel personalized federated learning (PFL) framework. We report on computational experiments using large language models (LLMs) as simulated agents to approximate perception in text-based emotion prediction, with inter- and intra-agent agreement guiding personalization and generalization decisions. For conversational interactions, we introduce FedSession, a tiered PFL architecture comprising (1) local model personalization, (2) intermediate aggregation to capture session-level interpersonal dynamics, and (3) global model generalization across sessions. The analyses reveal statistically significant correlations supporting the use of agreement for guiding personalization and generalization. Our findings reveal the limitations of prior PFL approaches in modeling heterogeneous behaviors and show that incorporating dyadic interaction through intermediate aggregation in FedSession improves performance over standard PFL techniques and local training baselines.
BIO: Rajesh Titung is a PhD student in Computing and Information Sciences in the Computational Linguistics and Speech Processing lab at RIT. His research interests include natural language processing and multimodality, human-in-the-loop machine learning, and affective AI. He focuses on developing novel insights about interactive machine learning and federated learning for multimodal and personalized computing. At RIT, he also participates in the AWARE-AI program and has conducted a research mobility visit at UCD/ML-Labs in Dublin, Ireland. Rajesh has co-mentored other students and has served as a teaching assistant for AI and natural language processing courses. Previously, he worked as a machine learning engineer in the industry.
NOTE: To schedule interpreter and/or services for this event, please use https://myaccess.rit.edu.
Event Snapshot
When and Where
Who
Open to the Public
| Cost | FREE |
Interpreter Requested?
No