2 Graphical Models in Action

  • From course Probabilistic Models and Inference Algorithms for Machine Learning, Prof. Dahua Lin
  • All contents here are from the course and self understandings.

基本步驟

  1. 理解你要求解的某個問題
    • what kind of entities/factors are involved?
    • How do they interact with each other?
    • Any constraints to take into account?
  2. 建立模型
    • Introduce variables
    • Specify relations among them: make assumptions and modeling choices
    • Formalize the graphical model
  3. Derive the inference & estimation algorithms

Gaussian Mixture Model (GMM)

1. Motivation & Assumptions

Gaussian Mixture Model

  • Observation: clusters (如圖中所示)
  • Assumptions:
    • latent components
    • independent generation of components (一個component指一個如圖所示的cluster)
    • independent generation of points in each component.

2. Formulate Gaussian Mixture Model

  • Variables:
    • sample points: xi
    • component indicators: zi
  • Generative procedure: for each i
    • choose a component: ziπ
    • generate a point: xiN(μzi,Σzi) (每個sample point 服從正態分佈,mean和方差爲μziΣzi )
  • Model parameters:
    • component parameters: {(μk,Σk)}k=1:K
    • choice prior: π=(π2,,πK)
  • Joint distribution:
    p(X,Z|Θ)=i=1NpC(zi|π)pN(xi|μzi,Σzi)
  • 該模型的圖形表達如下:
    Graphical Representation of GMM
    • 但是以上的模型不能泛化:當有多個groups 的數據時候,每一個 group Gm 都有一個 prior: πm ,這樣每一個group一個先驗,對於新來的數據,並沒有一個generalized的 πnew . 因此需要一個 Group-wise 的GMM

3. Extend Gaussian Mixture Model

  • 一般來說,要泛華某個先驗(prior),我們可以採用一些常見的分佈,如下,這裏我們採用Dirichlet
    Common Choices of Priors
  • Group-wise GMM: Generalizable to New Groups
    • Introduce a Dirichlet Prior over πm to allow the generation of new groups.
    • Formulation:
      • For each group Gm:πmDir(α) .
      • Generate the i-th point in Gm :
        ziπm
        xiN(μk,Σk)
      • 注意 πm 現在是個 隱變量 (latent variable)
        Group-wise GMM
  • Temporal Structures:
    • 在實際世界中,時域上的變化是很常見的,那對於模型來說,也需要對應的dynamics作用到 不同的變量上
    • Three ways to incorporate dynamics:
      Dynamics on xi
      Dynamics on zi
      Dynamics on π
發佈了57 篇原創文章 · 獲贊 3 · 訪問量 2萬+
發表評論
所有評論
還沒有人評論,想成為第一個評論的人麼? 請在上方評論欄輸入並且點擊發布.
相關文章