[Image Interpretation] Unsupervised Learning

Unsupervised Learning

Motivation: Before, some assumptions we adopted:
 1. a function that maps the the observed input to output;
 2. a training dataset to train the parameter for the model;
Without these assumption, we have
 1. unknown functional form: non-parametric density estimation: only get the probability value, without knowing the exact form of PDF(probability density function);
 2. only data, no output, data clustering.

Density Estimation

• Histogram: discretize the feature space into bins and count
  
  • Pro:
    
    • with infinite amount of data any density can be approximated arbitrarily well >> approach continuous
    • computationally simple
  • Con:
    
    • curse of dimensionality: the number of bins, and thus the number of data increase exponentially with the increase of data dimensionality;
    • the size of bins is hard to determine, and even no optimal size.

• Kernel density estimation

  • we are given an data point x first, then we need to output its probability Pr(x);
    $$Pr(x) = \frac{K}{NV} = \frac{1}{Nh^d}\sum^N_{i = 1} k_h(\textbf{x} -\textbf{x}_i)$$
  • where:
    
    • $h^d$ is the volume in d-dimensionality space;
    • N is the total number of given dataset
    • $k_h$ is the kernel function, where h stands for the kernel width;

  • Understanding:

    In general way, the kernel function $k_h(x - x_i)$ defines a weight function, which depends on the distance to examined location in feature space.
    While the denominator $NV$ can account for the terminology density, where over them accounts for the calculation of density.

  • Different kernel function:

    • Parzen window estimator:
• Bias-variance trade-off: refers to the influence radius of a point
  
  • For example, bin width of histogram, kernel width of kernel function, number of neighbors for kNN;
  • Too large radius result in too smooth case >> has large bias: a multimodal is mistakenly fitted into a single peak Gaussian
  • Too small radius result in too variant case >> has large variance: a single peak Gaussian is mistakenly fitted into a multimodal function;

Mixture Model

• Parameters we want to get in Expectation maximization(EM)
  
  • the mean $\mu_i$, vcm $\sum_i$ and occurrence probability (or called mixing coefficient for the i-th Gaussian distribution) $w_i$;
  • The updation of $\mu_i$, vcm $\sum_i$ and mixing coefficient $w_i$ required posterior probability of a data point $x_n$, after we initialize the three parameter: $\gamma(z_{nk}) = \sum_j \pi_j \mathcal{N}(\textbf{x}_n | \mu_j, \Sigma_j)$
  • See PRML at page 435.
• EM for change detection in video:
  
  • For each pixel we calculate a Gaussian Mixture function
  • then each time the newly generated pixel(from video flow) is put into the mixture model to calculate a probability, and check whether it is below a predefined threshold.
  • If the calculated probability is too low, then this means the pixel intensity changes, otherwise it remain stable.

Clustering

• K-means clustering
  
  • assumption:
    
    • we know the number k of clusters
    • each cluster is defined by a cluster center
    • data points are assigned to the nearest cluster center
  • ISODATA clustering:
    
    • in each iteration allows the change of cluster number
    • splitting of clusters with large standard deviation
    • merging of clusters with small distance between centres dissolving clusters with too few points
• Mean shift
  
  • from each data point, do local gradient ascent on the density
  • call all data points that reach the same mode a cluster
  • How to do local gradient ascent:
    
    • place a circular window at the data point
    • compute (weighted) mean of all points within the window;
    • weights depend on kernel function $k_h(x-x_i)$
    • move the window to the mean
    • iterate to convergence
  • Example: Image segmentation
    
    • Trick to achieve compact clustering: add pixel coordinate in the feature vector $\textbf{x} = \{R,G,B,X,Y\}$

Graph-based Clustering

• Agglomerative Clustering: Let each point be its own cluster, iteratively find the two most similar clusters and merge;
  
  • Different way to define distance:
    
    • single linkage: based on single point pair, not so robust; the cluster boundary along the low density valley, and cluster shape is irregular.
    • complete linkage: aim to keep group contact, cluster shape regular
• Divisive Clustering: let all points be one cluster, iteratively find the two most dissimilar one and divide it.
• Spectral Clustering: emphasize connectivity rather than compactness.
  
  • Motivation: Viewing the problem of clustering, or image segmentation, as problem of Minimum Graph Cut.
  • The elements of a graph including vertex, edge and weight of edge. So each data point in clustering, or each pixel in image, corresponding to vertex in graph. The edge and weight is a measurement of similarity between data point.
  • For the example of image segmentation, the weight can be calculated from Euclidean distance of feature vector for each pixel $\textbf{x} = \{R,G,B,X,Y\}$, using the following formula $w_{ij} = exp\{ -\frac{||x_i - x_j||}{2\sigma_2}\}$
  • Optimization goal: To find the cut edge whose sum of weight is the least: $minimize\, cut(A, B) = \sum_{i \in A} \sum_{j \in B} w_{ij}$;
  • Without adding constraint the above method tend to be biased >> generating isolated cluster.
  • To avoid this, we need to control the size of cluster (not so small), so we can put the size of cluster at the denominator:
    $$Ncut = cut(A, B) * (\frac{1}{Vol(A)} + \frac{1}{Vol(B)})$$
    Vol(A): volume of graph A: $Vol(A) = \sum_{i \in A}d_i$
    $d_i$: degree of each vertex, $d_i = \sum_j w_{ij}$
  • Procedure to solve spectral clustering
    
    • calculate graph Laplacian matrix:
      $$L = D - W = diag\{d_1, d_2, \dots, d_n \} - \left[ \begin{matrix} w_{11} & w_{12} & \cdots & w_{1n} \\ w_{21} & w_{22} & \cdots & w_{2n} \\ \vdots & \vdots & \ddots & \vdots \\ w_{n1} & w_{n2} & \cdots & w_{nn} \\ \end{matrix}\right]$$
    • graph Laplacian matrix satisfy $f^T L\,f = \frac{1}{2}\sum^n_{i = 1} w_{ij}(f_i - f_j)^2$, f can be any arbitrary vector;
    • By setting f as specific vector, the following transformation can be done: $f^T L\,f = |V| * NCut(A, B)$. Here |V| is a constant.
    • According to Rayleigh Quoitent: the maximum and minimum of $R(L, f) = \frac{f^T L\,f }{f^T f }$ happens to equal to its maximum and minimum eigenvalue, meanwhile as $f^T f$ is a constant, so minimizing $R(L, f)$ equals to minimize $f^T L\,f = |V| * NCut(A, B)$.
    • However, as the minimum eigenvalue of L equals to 0, whose corresponding eigenvector doesn’t satisfy the required condition, we just find the second minimum eigenvector v (a n-dimension vector) whose element is either positive or negative.
    • The positive element corresponding one cluster, while negative element corresponding to another.
    • Generalization for 3 or more cluster: Choose k eigenvectors with k smallest eigenvalue. Reordering the eigenvector as a new $N* K$ matrix, then use K-means to find the k clusters >> clustering after transformation.

Example of clustering

• Unsupervised learning
  
  • Requirement: different clusters should be well-separated and enough data should be provided.
• Image segmentation:

Reference

漫步Clustering