Self-Supervised 3D Behavior Representation Learning Based on Homotopic Hyperbolic Embedding.

Area of Science:

• Computer Science
• Machine Learning
• Artificial Intelligence

Background:

• Behavior sequences exhibit complex spatio-temporal interactions and high-dimensional nonlinear structures.
• Learning 3D behavior representations typically requires supervised signals, which are often unavailable.
• Existing self-supervised methods struggle to capture joint features in traditional Euclidean spaces.

Purpose of the Study:

• To develop a self-supervised learning method for mining nonlinear relationships in behavior trajectories.
• To overcome limitations of Euclidean spaces in representing context joint features.
• To improve unsupervised learning of 3D behavior representations.

Main Methods:

• Proposed a self-supervised learning method based on hyperbolic embedding for behavior trajectories.
• Employed contrastive learning focusing on global features and discarding negative samples.
• Utilized hyperbolic space embedding and multi-layer perceptron for homotopic mapping.

Main Results:

• The hyperbolic embedding method effectively mines nonlinear relationships in behavior data.
• The framework avoids issues associated with pulling similar data apart in feature space.
• Achieved improved performance in unsupervised learning of behavior representations.

Conclusions:

• Hyperbolic embedding combined with contrastive learning offers a powerful approach for unsupervised behavior representation learning.
• The method leverages geometric properties of hyperbolic manifolds and homotopy groups for enhanced learning.
• This approach provides better supervised signals for networks, advancing unsupervised learning capabilities.