- 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
- 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?
- 建立模型
- Introduce variables
- Specify relations among them: make assumptions and modeling choices
- Formalize the graphical model
- Derive the inference & estimation algorithms
Gaussian Mixture Model (GMM)
1. Motivation & Assumptions
- 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
- sample points:
- Generative procedure: for each i
- choose a component:
zi∈π - generate a point:
xi∈N(μzi,Σzi) (每個sample point 服從正態分佈,mean和方差爲μzi 和Σzi )
- choose a component:
- Model parameters:
- component parameters:
{(μk,Σk)}k=1:K - choice prior:
π=(π2,⋯,πK)
- component parameters:
- Joint distribution:
p(X,Z|Θ)=∏i=1NpC(zi|π)pN(xi|μzi,Σzi) - 該模型的圖形表達如下:
- 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
- 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
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
- 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:πm∼Dir(α) .
- For each group
- Generate the i-th point in
Gm :zi∼πm xi∼N(μk,Σk)
- Generate the i-th point in
- 注意
πm 現在是個 隱變量 (latent variable)
- 注意
- Introduce a Dirichlet Prior over
- 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 onxi
Dynamics onzi
Dynamics onπ