Clustering with k-means and Gaussian mixture distributions

Transcription

1 Clustering with k-means and Gaussian mixture distributions Machine Learning and Category Representation Jakob Verbeek, ovember 21, 2014 Course website:

2 Bag-of-words image representation in a nutshell 1) Sample local image patches, either using Interest point detectors (most useful for retrieval) Dense regular sampling grid (most useful for classification) 2) Compute descriptors of these regions For example SIFT descriptors 3) Aggregate the local descriptor statistics into global image representation This is where clustering techniques come in 4) Process images based on this representation Classification Retrieval

3 Bag-of-words image representation in a nutshell Frequency in image 3) Aggregate the local descriptor statistics into bag-of-word histogram Map each local descriptor to one of clusters (a.k.a. visual words ) Use -dimensional histogram of word counts to represent image Visual word index..

4 Example visual words found by clustering Airplanes Motorbikes Faces Wild Cats Leafs People Bikes

5 Clustering Finding a group structure in the data Data in one cluster similar to each other Data in different clusters dissimilar Maps each data point to a discrete cluster index in {1,..., } Flat methods do not suppose any structure among the clusters Hierarichal methods

6 Hierarchical Clustering Data set is organized into a tree structure Various level of granularity can be obtained by cutting-off the tree Top-down construction Start all data in one cluster: root node Apply flat clustering into groups Recursively cluster the data in each group Bottom-up construction Start with all points in separate cluster Recursively merge nearest clusters Distance between clusters A and B E.g. min, max, or mean distance between elements in A and B

7 Clustering descriptors into visual words Offline clustering: Find groups of similar local descriptors Using many descriptors from many training images Encoding a new image: Detect local regions Compute local descriptors Count descriptors in each cluster [5, 2, 3] [3, 6, 1]

8 Definition of k-means clustering Given: data set of points xn, n=1,, Goal: find cluster centers mk, k=1,, that minimize the squared distance to nearest cluster centers k k=1 E ({m } )= n=1 mink {1,..., } x n mk 2 Clustering = assignment of data points to nearest cluster center Indicator variables rnk=1 if xn assgined to mk, rnk=0 otherwise For fixed cluster centers, error criterion equals sum of squared distances between each data point and assigned cluster center k k=1 E ({m } )= n=1 k=1 r nk x n mk 2

9 Examples of k-means clustering Data uniformly sampled in unit square k-means with 5, 10, 15, and 25 centers

10 Minimizing the error function Goal find centers mk to minimize the error function k k =1 E ({m } )= n=1 mink {1,..., } x n m k 2 Any set of assignments, not necessarily the best assignment, gives an upper-bound on the error: k k=1 E({m } ) F ({mk }, {r nk })= n =1 k=1 r nk xn mk 2 The k-means algorithm iteratively minimizes this bound 1) Initialize cluster centers, eg. on randomly selected data points 2) Update assignments rnk for fixed centers mk 3) Update centers mk for fixed data assignments rnk 4) If cluster centers changed: return to step 2 5) Return cluster centers

11 Minimizing the error bound F ({mk },{r nk })= n=1 k=1 r nk x n mk 2 Update assignments rnk for fixed centers mk Constraint: exactly one rnk=1, rest zero k 2 r nk x n mk Decouples over the data points Solution: assign to closest center Update centers mk for fixed assignments rnk Decouples over the centers Set derivative to zero Put center at mean of assigned data points F =2 n r nk (x n mk )=0 mk n r nk x n mk = n r nk n r nk x n mk 2

12 Examples of k-means clustering Several k-means iterations with two centers Error function

13 Minimizing the error function k k=1 E ({m } )= n=1 mink {1,..., } x n mk 2 Goal find centers mk to minimize the error function Proceeded by iteratively minimizing the error bound k k=1 F ({m } )= n=1 k=1 r nk x n m k 2 -means iterations monotonically decrease error function since Both steps reduce the error bound Error bound matches true error after update of the assignments Bound #1 Bound #2 Minimum of bound #1 True error Error Placement of centers

14 Problems with k-means clustering Result depends heavily on initialization Run with different initializations eep result with lowest error

15 Problems with k-means clustering Assignment of data to clusters is only based on the distance to center o representation of the shape of the cluster Implicitly assumes spherical shape of clusters

16 Clustering with Gaussian mixture density Each cluster represented by Gaussian density Parameters: center m, covariance matrix C Covariance matrix encodes spread around center, can be interpreted as defining a non-isotropic distance around center Two Gaussians in 1 dimension A Gaussian in 2 dimensions

17 Clustering with Gaussian mixture density Each cluster represented by Gaussian density Parameters: center m, covariance matrix C Covariance matrix encodes spread around center, can be interpreted as defining a non-isotropic distance around center Definition of Gaussian density in d dimensions d /2 (x m, C )=(2π ) 1/2 C 1 T 1 exp (x m) C (x m) 2 Determinant of covariance matrix C ( ) Quadratic function of point x and mean m Mahanalobis distance

18 Mixture of Gaussian (MoG) density Mixture density is weighted sum of Gaussian densities Mixing weight: importance of each cluster p (x )= k =1 πk ( x mk, C k ) πk 0 Density has to integrate to 1, so we require k =1 πk =1 Mixture in 1 dimension Mixture in 2 dimensions What is wrong with this picture?!

19 Sampling data from a MoG distribution Let z indicate cluster index To sample both z and x from joint distribution p(z=k )=π k Select z with probability given by mixing weight Sample x from the z-th Gaussian p(x z=k )= ( x mk, C k ) MoG recovered if we marginalize over the unknown cluster index p(x )= k p( z=k) p(x z=k )= k π k (x mk, C k ) Color coded model and data of each cluster Mixture model and data from it

20 Soft assignment of data points to clusters Given data point x, infer cluster index z p(z=k, x) p(z=k x )= p(x ) π k (x mk, C k ) p (z=k ) p(x z=k ) = = k p(z=k ) p( x z=k) k π k (x m k,c k ) MoG model Data Color-coded soft-assignments

21 Clustering with Gaussian mixture density Given: data set of points xn, n=1,, Find mixture of Gaussians (MoG) that best explains data Maximize log-likelihood of fixed data set w.r.t. parameters of MoG Assume data points are drawn independently from MoG L(θ)= n=1 log p(x n ; θ) θ={π k, mk, C k }k=1 MoG learning very similar to k-means clustering Also an iterative algorithm to find parameters Also sensitive to initialization of paramters

22 Maximum likelihood estimation of single Gaussian Given data points xn, n=1,, Find single Gaussian that maximizes data log-likelihood d 1 1 T 1 L(θ)= n=1 log p( x n )= n=1 log (x n m, C)= n=1 log π log C (x n m) C ( x n m) ( Set derivative of data log-likelihood w.r.t. parameters to zero L(θ) =C 1 n=1 ( x n m ) =0 m 1 m= n=1 x n L(θ) 1 1 T = C (x n m)( xn m) =0 1 n=1 2 2 C 1 T C = n=1 ( x n m)( xn m) Parameters set as data covariance and mean ( ) )

23 Maximum likelihood estimation of MoG o simple equation as in the case of a single Gaussian Use EM algorithm Initialize MoG: parameters or soft-assign E-step: soft assign of data points to clusters M-step: update the mixture parameters Repeat EM steps, terminate if converged Convergence of parameters or assignments E-step: compute soft-assignments: qnk = p( z=k xn ) M-step: update Gaussians from weighted data points 1 πk = n=1 qnk 1 mk = q nk x n n=1 πk 1 T Ck= q (x m )( x m ) k n k πk n=1 nk n

24 Maximum likelihood estimation of MoG Example of several EM iterations

25 EM algorithm as iterative bound optimization Just like k-means, EM algorithm is an iterative bound optimization algorithm Goal: Maximize data log-likelihood, can not be done in closed form L(θ)= n =1 log p(x n )= n=1 log k=1 π k (x n mk, C k ) Solution: iteratively maximize (easier) bound on the log-likelihood Bound uses two information theoretic quantities Entropy ullback-leibler divergence

26 Entropy of a distribution Entropy captures uncertainty in a distribution Maximum for uniform distribution Minimum, zero, for delta peak on single value H (q)= k =1 q( z=k )log q( z=k ) Low entropy distribution High entropy distribution

27 Entropy of a distribution H (q)= k =1 q( z=k )log q( z=k ) Connection to information coding (oiseless coding theorem, Shannon 1948) Frequent messages short code, rare messages long code optimal code length is (at least) -log p bits Entropy: expected (optimal) code length per message Suppose uniform distribution over 8 outcomes: 3 bit code words Suppose distribution: 1/2,1/4, 1/8, 1/16, 1/64, 1/64, 1/64, 1/64, entropy 2 bits! Code words: 0, 10, 110, 1110, , ,111110, Codewords are self-delimiting : Do not need a space symbol to separate codewords in a string If first zero is encountered after 4 symbols or less, then stop. Otherwise, code is of length 6.

28 ullback-leibler divergence Asymmetric dissimilarity between distributions Minimum, zero, if distributions are equal Maximum, infinity, if p has a zero where q is non-zero q( z=k ) D (q p)= k =1 q( z=k ) log p( z=k ) Interpretation in coding theory Sub-optimality when messages distributed according to q, but coding with codeword lengths derived from p Difference of expected code lengths D (q p)= k=1 q (z=k )log p (z=k ) H (q ) 0 Suppose distribution q: 1/2,1/4, 1/8, 1/16, 1/64, 1/64, 1/64, 1/64 Coding with p: uniform over the 8 outcomes Expected code length using p: 3 bits Optimal expected code length, entropy H(q) = 2 bits L divergence D(q p) = 1 bit

29 EM bound on MoG log-likelihood We want to bound the log-likelihood of a Gaussian mixture p( x)= k =1 π k (x ; mk, C k ) Bound log-likelihood by subtracting L divergence D(q(z) p(z x)) Inequality follows immediately from non-negativity of L F (θ, q)=log p ( x ;θ) D ( q(z) p( z x,θ) ) log p( x ; θ) p(z x) true posterior distribution on cluster assignment q(z) an arbitrary distribution over cluster assignment Sum per data point bounds to bound the log-likelihood of a data set: F (θ, {q n })= n=1 log p(x n ;θ) D ( qn ( z) p( z x n,θ) ) n=1 log p (x n ; θ)

30 Maximizing the EM bound on log-likelihood E-step: fix model parameters, update distributions qn to maximize the bound F (θ, {q n })= n=1 [ log p (x n ) D (q n ( z n ) p (z n x n )) ] L divergence zero if distributions are equal Thus set qn(zn) = p(zn xn) After updating the qn the bound equals the true log-likelihood

31 Maximizing the EM bound on log-likelihood M-step: fix the soft-assignments qn, update model parameters F (θ, {q n })= n=1 [ log p (x n ) D (q n ( z n ) p ( z n x n )) ] [ = n=1 log p ( x n ) k q nk ( log qnk log p (z n =k x n ) ) [ = [ H (q )+ = n=1 H (q n )+ k q nk log p (z n =k, xn ) n n=1 k ] q nk ( log π k + log (x n ;m k,c k ) ) ] ] = k =1 n=1 qnk ( log πk +log ( x n ; m k, C k )) + n=1 H (q n ) Terms for each Gaussian decoupled from rest!

32 Maximizing the EM bound on log-likelihood Derive the optimal values for the mixing weights Maximize n=1 k =1 q nk log π k Take into account that weights sum to one, define Set derivative for mixing weight j >1 to zero n=1 qnj n=1 q n1 q log π k= π j π1 =0 π j n=1 k =1 nk n=1 qnj n=1 q n1 πj = π1 π 1 n=1 qnj=π j n=1 qn1 π 1 n=1 j=1 q nj = j=1 π j n qn1 π 1 = n=1 qn1 1 π j= n=1 qnj π1=1 k =2 πk

33 Maximizing the EM bound on log-likelihood Derive the optimal values for the MoG parameters For each Gaussian maximize n q nk log ( x n ; m k,c k ) Compute gradients and set to zero to find optimal parameters d 1 1 log ( x ; m,c )= log(2 π) log C ( x n m)t C 1 ( x n m) log ( x ; m,c )=C 1 ( x m) m log ( x ; m,c )= 1 C 1 ( x m)( x m)t 2 2 C 1 n q nk x n mk = n q nk n q nk ( x n m)( x n m)t C k= n qnk

34 EM bound on log-likelihood L is bound on data log-likelihood for any distribution q F (θ, {q n })= n=1 [ log p (x n ) D (q n ( z n ) p (z n x n )) ] F (θ, {q n }) Iterative coordinate ascent on F E-step optimize q, makes bound tight M-step optimize parameters F (θ, {q n }) F (θ, {q n }) F (θ, {q n })

35 Clustering with k-means and MoG Assignment: -means: hard assignment, discontinuity at cluster border MoG: soft assignment, 50/50 assignment at midpoint Cluster representation -means: center only MoG: center, covariance matrix, mixing weight If mixing weights are equal and all covariance matrices are constrained to be then EM algorithm = k-means algorithm C k=ϵ I and ϵ 0 For both k-means and MoG clustering umber of clusters needs to be fixed in advance Results depend on initialization, no optimal learning algorithms Can be generalized to other types of distances or densities

36 Reading material More details on k-means and mixture of Gaussian learning with EM Pattern Recognition and Machine Learning, Chapter 9 Chris Bishop, 2006, Springer