RBM with Local Interactions

Method

Goal: model high-dimentional motion data.

Motivation: conventional restricted Boltzmann machine treats each input as a scalar. When handling sequential data, each dimension is modeled individually. We would like to model the spatial interations in the vector data.

LRBM

Proposed methods

  • Learning: extend the conventional contrastive divergence for parameter learning.
  • Inference: estimate the relative partition function for likelihood inference and classification.
  • Applications: we evaluate the learning and inference algorithm on two computer vision tasks involving high-dimensional motion data: facial expression recognition and human action recognition.

    • Publications

    Articles in journals

    Siqi Nie, Ziheng Wang, and Qiang Ji, "A Generative Restricted Boltzmann Machine based Method for High-Dimensional Motion Data Modeling," Computer Vision and Image Understanding (CVIU), Vol. 136, pp. 14-22, 2015. [link]

    Articles in conference proceedings

    Siqi Nie and Qiang Ji, "Capturing Global and Local Dynamics for Human Action Recognition," in Proceedings of the 22nd International Conference on Pattern Recognition (ICPR), 2014. Oral Presentation. [PDF]