RVM Implementation Progress

1
RVM Implementation Progress
Seungchan Lee Department of Electrical and
Computer Engineering Mississippi State University
2

• Relevance Vector Machine

• Why is Relevance Vector Machine?
• Problems of SVM
• Use large number of basis function
• Binary classification, (does not consider
  probabilistic methods )
• How can overcome this?
• Adopt probabilistic Bayesian approach
• Relevance vector learning becomes the
  optimization of hyperparameters.

3

• Progress

• Current problem
• Initialization problem for large data set
• Inverse matrix problem
• RVM class (in progress)
• Progress as of (Wons work)
• Positive definite matrix problem
• The error occurs on Cholesky decomposition step
• In order to fix this problem, it is required to
  fully understand RVM algorithm.
• Make simple IFC program
• 300 data samples in-class and out-class started
  to cause the positive definite problem

4

• Plan

• Fix positive definite matrix problem in the
  Cholesky decomposition step
• check the positive definite matrix before this
  step
• Improving efficiency
• Divide and Conquer Approach
• To make a whole vector to small subset of vectors
• Find alternatives speech recognition needs
  larger size of vectors
• Reduce memory and computation time of the
  Cholesky decomposition step
• Constructive approach by Tipping and Faul
• Sequential Bootstrapped SVM method