Presented at the International Conference on Animal-Computer Interaction, this study by Devon Martin, David L. Roberts, and Alper Bozkurt addresses a fundamental challenge in guide dog programs: identifying behavioral and temperamental traits that predict successful completion of training. Currently, a major “blind spot” in evaluation occurs during the period when puppies are raised by volunteers, where less structured observation limits predictive accuracy.
The researchers used a custom-designed smart collar to gather continuous environmental and behavioral data from puppies as they progressed through various stages of guide dog socialization and training. This real-world dataset forms a critical contrast to traditional manifold learning experiments, which typically rely on controlled laboratory settings and thus do not translate well to operational environments.
To classify behaviors and extract meaningful patterns, the team evaluated three machine learning approaches:
• Long short-term memory networks (LSTMs) for modeling temporal dependencies in behavioral sequences.
• Autoencoders (AEs) for unsupervised feature extraction and dimensionality reduction.
• Kernel principal component analysis (KPCA) for nonlinear manifold learning.
They additionally tested multi-sensor data fusion to determine the most informative combinations of sensor modalities—such as motion, environmental cues, or physiological signals—for classifying temperament-related behaviors.
The overarching goal is to develop a data-pattern-to-behavior dictionary capable of supporting large-scale, in-the-field canine behavior modeling. By leveraging In For Training (IFT) program data—collected in minimally controlled environments—the researchers aim to bridge the gap between experimental machine learning applications and practical operational use in guide dog programs.
This work highlights the emerging potential of wearable canine technology combined with advanced behavioral modeling to improve selection, training, and long-term success of guide dog candidates.
Source: Martin, D., Roberts, D. L., & Bozkurt, A. (2023). Toward In-the-Field Canine Manifold Learning: Data Fusion for Evaluation of Potential Guide Dogs. Presented December 4, 2023.
