|Title: Learning Movement Representations of Small Humans with Small Data
|Seminar: Computer Science
|Speaker: Sarah Ostadabbas, Northeastern University
|Contact: Vaidy Sunderam, firstname.lastname@example.org
|Date: 2022-11-04 at 1:00PM
|Venue: MSC W201
|Download Flyer Add to Calendar
Closely tracking the development of motor functioning in infants provides prodromal risk markers of many developmental disruption such as autism spectrum disorder (ASD), cerebral palsy (CP), and developmental coordination disorder (DCD), among others. Screening for motor delays will allow for earlier and more targeted interventions that will have a cascading effect on multiple domains of infant development, including communication, social, cognitive, and memory. However, only about 29% of US children under 5 years of age receive developmental screening due to expense and shortage of testing resources, contributing negatively to lifelong outcomes for infants at risk for developmental delays. My research aims to learn and quantify visual representations of motor function in infants towards designing an accessible and affordable video-based screening technology for their motor skills by developing novel data-/label-efficient AI techniques including biomechanically-constrained synthetic data augmentation, semantic-aware domain adaptation, and human-AI co-labeling algorithms.
While there are several powerful human behavior recognition and tracking algorithms, however, models trained on large-scale adult activity datasets have limited success in estimating infant movements due to the significant differences in their body ratios, the complexity of infant poses, and types of their activities. Privacy and security considerations hinder the availability of adequate infant images/videos required for training of a robust model with deep structure from scratch, making this a particularly constrained ``small data problem''. To address this gap, in this talk I will cover: (i) introduction of biomechanically-constrained models to synthesize labeled pose data in the form of domain-adjacent data augmentation; (ii) design and analysis of a semantic-aware unsupervised domain adaptation technique to close the gap between the domain-adjacent and domain-specific pose data distributions; and (iii) development and analysis of an AI-human co-labeling technique to provide high-quality labels to refine and adapt the domain-adapted inference models into robust pose estimation algorithms in the target application. These contributions enable the use of advanced AI in the small data domain.
Zoom Option: https://emory.zoom.us/j/95719302738
See All Seminars