Semester of Graduation

Fall 2025

Degree Type

Dissertation/Thesis

Degree Name

Masters of Science in Artificial Intelligence

Department

Computer Science

Committee Chair/First Advisor

Dr. Zongxing Xie

Second Advisor

Dr. Honghui Xu

Third Advisor

Dr. Xinyue Zhang

Abstract

As neurodegenerative diseases increase in prevalence, there is an increased need for early indicators such as gait-based markers of cognitive decline. Although gait markers predict future degradation, most evidence comes from controlled laboratory settings and does not generalize well to free-living environments, reducing its applicability to asymptomatic individuals. Free-living gait introduces uncontrolled factors including cognitive load from dual-task behaviors, which alters gait features in ways that are difficult to interpret in some modalities such as IMU, as the resultant kinematic deviations are much more subtle than those between broad activity categories.

We present a cross-modal adaptation framework that leverages paired data to transfer knowledge from a more informative modality (MoCap) to a less informative modality (IMU) to better learn task-relevant representations for dual-task activity classification. The framework integrates supervised and self-supervised contrastive alignment with multi-task learning and is evaluated across four encoder types and three cross-modal strategies. We find that cross-modal contrastive learning improves self-supervised TCN classification by 6.5%, supervised contrastive pretraining yields improvement of up to 13%, and auxiliary contextual signals such as kinematic regression and adversarial identity suppression enhance representation quality in some low-supervision scenarios.

These findings demonstrate that cross-modal contrastive supervision and multi-task learning markedly improve dual-task gait recognition from low-information modalities lacking large quantities of annotated data. This is significant as these techniques can be used to mitigate the need to produce those annotations, significantly reducing the cost of deployment of free-living gait analysis for early cognitive assessment.

Comments

This work is supported by the Interdisciplinary Seed Grant from the KSU Office of Research

defense form.pdf (212 kB)

Available for download on Wednesday, December 16, 2026

Share

COinS