| MS Seminar

Name of the Speaker: Mr. Vishnu Nair (EE21S064)
Guide: Dr. Mohanasankar S
Co-Guide: Dr. Jayaraj Joseph
Online meeting link: https://meet.google.com/ywt-snww-twc
Date/Time: 6th May 2025 (Tuesday), 4:30 PM
Title: Implicit and Explicit Compositionality in Spatio-Temporal Data Models.

Abstract :

This seminar presents research on compositionality in spatio-temporal data models, exploring how a finite set of fundamental primitives can compose to form complex structures. The study approaches this problem in two distinct ways. First, it examines the implicit compositionality emerging in attention-only transformers trained on spatio-temporal data. Using Mechanistic Interpretability methods, the internal representations and compositional structure of these models are analyzed and quantified. A novel toy dataset, "Moving Dots," is developed to support this investigation, enabling controlled experimentation with spatio-temporal patterns. Second, explicit compositional models are built using Symbolic Programming, where shape-based landmark trajectories—segmented using change detection techniques—are used as primitives. The primary spatio-temporal data employed is skeleton-based action data. Compositional representations are shown to be effective for such data. Evidence for inherent compositionality in transformer models trained on such data is also presented.