Cross-Species Behavioral Representation Learning Using Domain-Adversarial Adaptation on Wearable IMU Signals

Area of Science:

• Animal behavior analysis
• Machine learning for biomechanics
• Wearable sensor technology

Background:

• Current animal activity recognition systems using inertial measurement unit (IMU) sensors and deep learning are often species-dependent, requiring extensive labeled data for each new species.
• Significant biomechanical and anatomical differences between species pose challenges for generalizable behavior recognition.

Purpose of the Study:

• To develop a biomimetic, cross-species framework for learning transferable locomotor structures from heterogeneous IMU signals.
• To reduce inter-species distributional discrepancies in animal behavior recognition.

Main Methods:

• A novel framework integrating behavioral ontology harmonization, imbalance-aware augmentation, and semi-supervised domain-adversarial adaptation.
• A dual-head architecture for simultaneous processing of multi-label and single-label behavioral structures across diverse species.
• Experiments on dog, goat, and horse datasets to evaluate cross-species transferability.

Main Results:

• The proposed framework significantly improved cross-species transferability, achieving a mean Macro-F1 score of 0.711 compared to 0.345 for direct transfer learning.
• Sparse target supervision was found critical for stabilizing adversarial adaptation.
• Semi-supervised Domain-Adversarial Neural Network (DANN) outperformed K-Means head adaptation, highlighting the need for sparse target supervision.

Conclusions:

• Biologically motivated locomotor similarity can facilitate cross-species behavioral transfer in animal activity recognition.
• The developed framework offers a promising approach for more generalizable animal behavior analysis using wearable sensors.
• Further validation across more species, sensor placements, and deployment conditions is recommended.