Skip to content

qsh-zh/cbp

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

89 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

CBP(Constrained Belief Propagation)

Constrained Belief Progapation is a package for solving inference tasks with collective\aggregate evidence.

What is collective\aggregate evidence?

Let us consider a task, estimating dymanics of bird migration. We consider the following different kinds of evidence

  1. Install trackers for interested birds. So we access to the trajectories of sampled birds. We can query the position of interested birds at any time and how they move.

  2. We shoot a video for the birds migration. We can quety the distibution of the birds in the space at any time. However, we do not have access to individual trajectories and how birds move overtime.

The second information is an instance of collective\aggregate evidence, which remove the distinguishability of individuals.

Why do we need collective\aggregate evidence?

  • Easy and cheap to acquire

  • Privacy concern

Sometimes, collective\aggregate evidence is the ony information we access to analysis collective behavior.

How can we represent the special evidence?

In CBP, we represent the special evidence as a special nodes in probabilistic graphical model(PGM), named node with fixed marginal constraints. A hidden markov model with aggregate evidence can be presented as:

image-20200716201949266

The shaded nodes represents the aggregate evidence. We can introduce a simple version:

image-20200716202622245

How to do inference?

In CBP framework, we present a nice elegant connections between multi-marginal optimal transport problems and inference problem for PGM. We implement algorithms similar to the standard belief propagation. In CBP, we present Iterative Scaling Belief Propagation and Contrained Norm-product two algorithms. More details can be find the paper.

Citation

@article{haasler2020multi,
  title={Multi-marginal optimal transport and probabilistic graphical models},
  author={Haasler, Isabel and Singh, Rahul and Zhang, Qinsheng and Karlsson, Johan and Chen, Yongxin},
  journal={arXiv preprint arXiv:2006.14113},
  year={2020}
}