Meta-Learning Coordinated Eye–Hand Dynamics

The analysis of fine-grained behavioral time-series, such as eye-hand coordination, offers a powerful lens for understanding neurodevelopmental disorders like Autism Spectrum Disorder (ASD). Traditional assessments often miss subtle dynamic patterns that are critical for characterizing individual differences. This thesis investigates the temporal dynamics of coordinated eye and hand behavior in adolescents with and without ASD during a structured ”maze painting” tablet task. The goal is not to propose a new predictive model, but to use a state-of-the-art meta-learning framework based on latent Neural Ordinary Differential Equations (Neural ODEs) as a ”computational microscope” to explore how these behavioral dynamics vary within and across individuals and task contexts. By mapping high-dimensional, irregular eye-gaze and touch data into a low-dimensional latent space, we analyze the resulting trajectories to uncover patterns related to individual identity and temporal evolution. The results, visualized using t-SNE, demonstrate that the model successfully learns to encode stable, individual-specific behavioral signatures. Notably, the clarity of these individual signatures is timescale-dependent, becoming significantly more pronounced in longer-duration windows, which suggests that unique behavioral traits emerge over multi-second intervals. This work validates the use of la- tent dynamics models for scientific inquiry, revealing nuanced, person-specific patterns in sensorimotor behavior that could inform future digital biomarker development for ASD.