GMM

GMM

Cluster assignment of point

• $P(Z=j)$ describe by $P(C_j) = w_j \geq 0$ with $\sum_{j=1}^k w_j = 1$

Each cluster has its own Gaussian distribution

• $P(X|Z=j) = \mathcal{N}(\mu_j, \Sigma_j)$ class conditional density

Gaussian mixture distribution:

$$P(x) \equiv \sum_{j=1}^k w_j \mathcal{N}(x|\mu_j, \Sigma_j)$$

Modelling the data points as independent, aim is to find $P(C_j), \mu_j, \Sigma_j, j = 1,\ldots, k$ that maximise

$P(x_1, \dots, x_n) = \prod_{i=1}^n \sum_{j=1}^k P(C_j) P(x_i | C_j)$

where $P(x | C_j) \equiv \mathcal{N}(x | \mu_j, \Sigma_j)$

Try the usual log trick

$$\log P(x_1, \dots, x_n) = \sum_{i=1}^n \log \left( \sum_{j=1}^k P(C_j) P(x_i | C_j) \right)$$

EM

MLE is a frequentist principle that suggests that given a daraset, the "best" parameters to use are the ones that maximise the probability of the data

• MLE is a way to formally pose the problem

EM is an algorihm

• EM is a way to solve the problem posed by MLE
• Especially convenient under unvserved data MLE can be found by other methods such as gradient descent

EM is a general approach, goes beyond GMMs

• Purpose: Implement MLE underlatent variables Z

Variableasz in GMMs:

• Variables: Point locations X abd cluster assignments Z
• Parameters : $\theta$ are cluster locations and scales

What is EM really doing?

• Coordinate ascent on a lower bound on the log-likelihood

  • M-step: ascent in modelede parameters $\theta$
  • E-step: ascent in the marginal likelihood $P(z)$

• Each step moves towards a local optimum

• Can get stuck, can need random restarts

EM For fitting the GMM

Initialisation Step:

• Initialze K clusters $C1, \ldots, C_k (\mu_j, \Sigma_j)$ and $P(C_j)$ for each cluster j.

Iterationm Step:

• Estimate the cluster of each datum $p(C_j | x_j) \rightarrow$ Expection
• Re-estimate the cluster parameters $(\mu_j, \Sigma_j), p(C_j)$ for each cluster j