HVGH: Unsupervised Segmentation for High-Dimensional Time Series Using Deep Neural Compression and Statistical...

Area of Science:

• Robotics
• Machine Learning
• Artificial Intelligence

Background:

• Humans segment continuous high-dimensional information into meaningful units for perception and learning.
• Unsupervised segmentation is vital for robots to learn complex tasks like language and motion.
• Previous models like HDP-GP-HSMM struggle with high-dimensional data and require pre-extracted features.

Purpose of the Study:

• To develop a novel model capable of unsupervised feature extraction and segmentation for high-dimensional time-series data.
• To overcome the limitations of previous models in handling complex, high-dimensional data for robotic learning.
• To enable robots to learn from raw, high-dimensional data without manual feature engineering.

Main Methods:

• Proposed a hierarchical Dirichlet process-variational autoencoder-Gaussian process-hidden semi-Markov model (HVGH).
• Employed a mutual learning loop between a variational autoencoder and the HDP-GP-HSMM for parameter estimation.
• Integrated feature extraction and segmentation within a single unsupervised framework.

Main Results:

• The HVGH model successfully extracts features from high-dimensional time-series data.
• The model performs unsupervised segmentation of the data.
• Experiments with motion-capture data demonstrated accurate segmentation and correct class number estimation, outperforming baseline methods.
• The method learned a latent space suitable for segmentation.

Conclusions:

• The proposed HVGH model offers a significant advancement in unsupervised learning for high-dimensional time-series data.
• This approach enables robots to learn complex patterns in data without manual feature engineering.
• HVGH provides a robust framework for segmentation and feature learning in robotics and AI.