Latent Variable Models with Gaussian Processes

Transcription

1 Latent Variable Models with Gaussian Processes Neil D. Lawrence GP Master Class 6th February 2017

2 Outline Motivating Example Linear Dimensionality Reduction Non-linear Dimensionality Reduction

3 Outline Motivating Example Linear Dimensionality Reduction Non-linear Dimensionality Reduction

4 Motivation for Non-Linear Dimensionality Reduction USPS Data Set Handwritten Digit 3648 Dimensions 64 rows by 57 columns

5 Motivation for Non-Linear Dimensionality Reduction USPS Data Set Handwritten Digit 3648 Dimensions 64 rows by 57 columns Space contains more than just this digit.

6 Motivation for Non-Linear Dimensionality Reduction USPS Data Set Handwritten Digit 3648 Dimensions 64 rows by 57 columns Space contains more than just this digit. Even if we sample every nanosecond from now until the end of the universe, you won t see the original six!

7 Motivation for Non-Linear Dimensionality Reduction USPS Data Set Handwritten Digit 3648 Dimensions 64 rows by 57 columns Space contains more than just this digit. Even if we sample every nanosecond from now until the end of the universe, you won t see the original six!

8 Simple Model of Digit Rotate a Prototype

9 Simple Model of Digit Rotate a Prototype

10 Simple Model of Digit Rotate a Prototype

11 Simple Model of Digit Rotate a Prototype

12 Simple Model of Digit Rotate a Prototype

13 Simple Model of Digit Rotate a Prototype

14 Simple Model of Digit Rotate a Prototype

15 Simple Model of Digit Rotate a Prototype

16 Simple Model of Digit Rotate a Prototype

17 MATLAB Demo demdigitsmanifold([1 2], all )

18 MATLAB Demo demdigitsmanifold([1 2], all ) PC no PC no 1

19 MATLAB Demo demdigitsmanifold([1 2], sixnine ) PC no PC no 1

20 Low Dimensional Manifolds Pure Rotation is too Simple In practice the data may undergo several distortions. e.g. digits undergo thinning, translation and rotation. For data with structure : we expect fewer distortions than dimensions; we therefore expect the data to live on a lower dimensional manifold. Conclusion: deal with high dimensional data by looking for lower dimensional non-linear embedding.

21 Outline Motivating Example Linear Dimensionality Reduction Non-linear Dimensionality Reduction

22 Notation q dimension of latent/embedded space p dimension of data space n number of data points data, Y = [ y 1,:,..., y n,: ] = [ y :,1,..., y :,p ] R n p centred data, Ŷ = [ ŷ 1,:,..., ŷ n,: ] = [ ŷ :,1,..., ŷ :,p ] R n p, ŷ i,: = y i,: µ latent variables, X = [ x 1,:,..., x n,: ] = [ x :,1,..., x :,q ] R n q mapping matrix, W R p q a i,: is a vector from the ith row of a given matrix A a :,j is a vector from the jth row of a given matrix A

23 Reading Notation X and Y are design matrices Data covariance given by 1 nŷ Ŷ cov (Y) = 1 n n ŷ i,: ŷ i,: = nŷ 1 Ŷ = S. i=1 Inner product matrix given by YY K = ( k i,j )i,j, k i,j = y i,: y j,:

24 Linear Dimensionality Reduction Find a lower dimensional plane embedded in a higher dimensional space. The plane is described by the matrix W R p q. x2 y = Wx + µ x 1 y 2 y 3 y 1 Figure: Mapping a two dimensional plane to a higher dimensional space in a linear way. Data are generated by corrupting points on the plane with noise.

25 Linear Dimensionality Reduction Linear Latent Variable Model Represent data, Y, with a lower dimensional set of latent variables X. Assume a linear relationship of the form y i,: = Wx i,: + ɛ i,:, where ɛ i,: N ( 0, σ 2 I ).

26 Linear Latent Variable Model Probabilistic PCA Define linear-gaussian relationship between latent variables and data. W Y X σ 2 n p (Y X, W) = N ( y i,: Wx i,:, σ 2 I ) i=1

27 Linear Latent Variable Model Probabilistic PCA Define linear-gaussian relationship between latent variables and data. Standard Latent variable approach: W Y X σ 2 n p (Y X, W) = N ( y i,: Wx i,:, σ 2 I ) i=1

28 Linear Latent Variable Model Probabilistic PCA Define linear-gaussian relationship between latent variables and data. Standard Latent variable approach: Define Gaussian prior over latent space, X. X W Y σ 2 n p (Y X, W) = N ( y i,: Wx i,:, σ 2 I ) i=1 n p (X) = N ( x i,: 0, I ) i=1

29 Linear Latent Variable Model Probabilistic PCA Define linear-gaussian relationship between latent variables and data. Standard Latent variable approach: Define Gaussian prior over latent space, X. Integrate out latent variables. X W Y σ 2 n p (Y X, W) = N ( y i,: Wx i,:, σ 2 I ) i=1 n p (X) = N ( x i,: 0, I ) i=1 n p (Y W) = N ( y i,: 0, WW + σ 2 I ) i=1

30 Computation of the Marginal Likelihood y i,: = Wx i,: + ɛ i,:, x i,: N (0, I), ɛ i,: N ( 0, σ 2 I )

31 Computation of the Marginal Likelihood y i,: = Wx i,: + ɛ i,:, x i,: N (0, I), ɛ i,: N ( 0, σ 2 I ) Wx i,: N ( 0, WW ),

32 Computation of the Marginal Likelihood y i,: = Wx i,: + ɛ i,:, x i,: N (0, I), ɛ i,: N ( 0, σ 2 I ) Wx i,: N ( 0, WW ), Wx i,: + ɛ i,: N ( 0, WW + σ 2 I )

33 Linear Latent Variable Model II Probabilistic PCA Max. Likelihood Soln (Tipping and Bishop, 1999) W Y σ 2 p (Y W) = n N ( y i,: 0, WW + σ 2 I ) i=1

34 Linear Latent Variable Model II Probabilistic PCA Max. Likelihood Soln (Tipping and Bishop, 1999) p (Y W) = n N ( y i,: 0, C ), i=1 C = WW + σ 2 I

35 Linear Latent Variable Model II Probabilistic PCA Max. Likelihood Soln (Tipping and Bishop, 1999) p (Y W) = n N ( y i,: 0, C ), i=1 C = WW + σ 2 I log p (Y W) = n 2 log C 1 2 tr ( C 1 Y Y ) + const.

36 Linear Latent Variable Model II Probabilistic PCA Max. Likelihood Soln (Tipping and Bishop, 1999) p (Y W) = n N ( y i,: 0, C ), i=1 C = WW + σ 2 I log p (Y W) = n 2 log C 1 2 tr ( C 1 Y Y ) + const. If U q are first q principal eigenvectors of n 1 Y Y and the corresponding eigenvalues are Λ q,

37 Linear Latent Variable Model II Probabilistic PCA Max. Likelihood Soln (Tipping and Bishop, 1999) p (Y W) = n N ( y i,: 0, C ), i=1 C = WW + σ 2 I log p (Y W) = n 2 log C 1 2 tr ( C 1 Y Y ) + const. If U q are first q principal eigenvectors of n 1 Y Y and the corresponding eigenvalues are Λ q, W = U q LR, L = ( Λ q σ 2 I ) 1 2 where R is an arbitrary rotation matrix.

38 Outline Motivating Example Linear Dimensionality Reduction Non-linear Dimensionality Reduction

39 Difficulty for Probabilistic Approaches Propagate a probability distribution through a non-linear mapping. Normalisation of distribution becomes intractable. x2 y j = f j (x) x 1 Figure: A three dimensional manifold formed by mapping from a two dimensional space to a three dimensional space.

40 Difficulty for Probabilistic Approaches x y 1 = f 1 (x) y 2 = f 2 (x) y2 y 1 Figure: A string in two dimensions, formed by mapping from one dimension, x, line to a two dimensional space, [y 1, y 2 ] using nonlinear functions f 1 ( ) and f 2 ( ).

41 Difficulty for Probabilistic Approaches y = f (x) + ɛ p(x) p(y) Figure: A Gaussian distribution propagated through a non-linear mapping. y i = f (x i ) + ɛ i. ɛ N ( 0, 0.2 2) and f ( ) uses RBF basis, 100 centres between -4 and 4 and l = 0.1. New distribution over y (right) is multimodal and difficult to normalize.

42 Linear Latent Variable Model III Dual Probabilistic PCA Define linear-gaussian relationship between latent variables and data. W Y X σ 2 n p (Y X, W) = N ( y i,: Wx i,:, σ 2 I ) i=1

43 Linear Latent Variable Model III Dual Probabilistic PCA Define linear-gaussian relationship between latent variables and data. Novel Latent variable approach: W Y X σ 2 n p (Y X, W) = N ( y i,: Wx i,:, σ 2 I ) i=1

44 Linear Latent Variable Model III Dual Probabilistic PCA Define linear-gaussian relationship between latent variables and data. Novel Latent variable approach: Define Gaussian prior over parameters, W. W X Y σ 2 n p (Y X, W) = N ( y i,: Wx i,:, σ 2 I ) i=1 p p (W) = N ( w i,: 0, I ) i=1

45 Linear Latent Variable Model III Dual Probabilistic PCA Define linear-gaussian relationship between latent variables and data. Novel Latent variable approach: Define Gaussian prior over parameters, W. Integrate out parameters. W X Y σ 2 n p (Y X, W) = N ( y i,: Wx i,:, σ 2 I ) i=1 p p (W) = N ( w i,: 0, I ) i=1 p p (Y X) = N ( y :,j 0, XX + σ 2 I ) j=1

46 Computation of the Marginal Likelihood y :,j = Xw :,j +ɛ :,j, w :,j N (0, I), ɛ i,: N ( 0, σ 2 I )

47 Computation of the Marginal Likelihood y :,j = Xw :,j +ɛ :,j, w :,j N (0, I), ɛ i,: N ( 0, σ 2 I ) Xw :,j N ( 0, XX ),

48 Computation of the Marginal Likelihood y :,j = Xw :,j +ɛ :,j, w :,j N (0, I), ɛ i,: N ( 0, σ 2 I ) Xw :,j N ( 0, XX ), Xw :,j + ɛ :,j N ( 0, XX + σ 2 I )

49 Linear Latent Variable Model IV Dual Probabilistic PCA Max. Likelihood Soln (Lawrence, 2004, 2005) X Y σ 2 p (Y X) = p N ( y :,j 0, XX + σ 2 I ) j=1

50 Linear Latent Variable Model IV Dual PPCA Max. Likelihood Soln (Lawrence, 2004, 2005) p (Y X) = p N ( y :,j 0, K ), j=1 K = XX + σ 2 I

51 Linear Latent Variable Model IV PPCA Max. Likelihood Soln (Tipping and Bishop, 1999) p (Y X) = p N ( y :,j 0, K ), j=1 K = XX + σ 2 I log p (Y X) = p 2 log K 1 2 tr ( K 1 YY ) + const.

52 Linear Latent Variable Model IV PPCA Max. Likelihood Soln p (Y X) = p N ( y :,j 0, K ), j=1 K = XX + σ 2 I log p (Y X) = p 2 log K 1 2 tr ( K 1 YY ) + const. If U q are first q principal eigenvectors of p 1 YY and the corresponding eigenvalues are Λ q,

53 Linear Latent Variable Model IV PPCA Max. Likelihood Soln p (Y X) = p N ( y :,j 0, K ), j=1 K = XX + σ 2 I log p (Y X) = p 2 log K 1 2 tr ( K 1 YY ) + const. If U q are first q principal eigenvectors of p 1 YY and the corresponding eigenvalues are Λ q, X = U qlr, L = ( Λ q σ 2 I ) 1 2 where R is an arbitrary rotation matrix.

54 Linear Latent Variable Model IV Dual PPCA Max. Likelihood Soln (Lawrence, 2004, 2005) p (Y X) = p N ( y :,j 0, K ), j=1 K = XX + σ 2 I log p (Y X) = p 2 log K 1 2 tr ( K 1 YY ) + const. If U q are first q principal eigenvectors of p 1 YY and the corresponding eigenvalues are Λ q, X = U qlr, L = ( Λ q σ 2 I ) 1 2 where R is an arbitrary rotation matrix.

55 Linear Latent Variable Model IV PPCA Max. Likelihood Soln (Tipping and Bishop, 1999) p (Y W) = n N ( y i,: 0, C ), i=1 C = WW + σ 2 I log p (Y W) = n 2 log C 1 2 tr ( C 1 Y Y ) + const. If U q are first q principal eigenvectors of n 1 Y Y and the corresponding eigenvalues are Λ q, W = U q LR, L = ( Λ q σ 2 I ) 1 2 where R is an arbitrary rotation matrix.

56 Equivalence of Formulations The Eigenvalue Problems are equivalent Solution for Probabilistic PCA (solves for the mapping) Y YU q = U q Λ q W = U q LR Solution for Dual Probabilistic PCA (solves for the latent positions) YY U q = U qλ q X = U qlr Equivalence is from U q = Y U qλ 1 2 q

57 Non-Linear Latent Variable Model Dual Probabilistic PCA Define linear-gaussian relationship between latent variables and data. Novel Latent variable approach: Define Gaussian prior over parameteters, W. Integrate out parameters. W X Y σ 2 n p (Y X, W) = N ( y i,: Wx i,:, σ 2 I ) i=1 p p (W) = N ( w i,: 0, I ) i=1 p p (Y X) = N ( y :,j 0, XX + σ 2 I ) j=1

58 Non-Linear Latent Variable Model Dual Probabilistic PCA Inspection of the marginal likelihood shows... W X Y σ 2 p p (Y X) = N ( y :,j 0, XX + σ 2 I ) j=1

59 Non-Linear Latent Variable Model Dual Probabilistic PCA Inspection of the marginal likelihood shows... The covariance matrix is a covariance function. W Y X σ 2 p p (Y X) = N ( y :,j 0, K ) j=1 K = XX + σ 2 I

60 Non-Linear Latent Variable Model Dual Probabilistic PCA Inspection of the marginal likelihood shows... The covariance matrix is a covariance function. We recognise it as the linear kernel. W X Y p p (Y X) = N ( y :,j 0, K ) j=1 K = XX + σ 2 I σ 2 This is a product of Gaussian processes with linear kernels.

61 Non-Linear Latent Variable Model Dual Probabilistic PCA Inspection of the marginal likelihood shows... The covariance matrix is a covariance function. We recognise it as the linear kernel. We call this the Gaussian Process Latent Variable model (GP-LVM). W X Y σ 2 p p (Y X) = N ( y :,j 0, K ) j=1 K =? Replace linear kernel with non-linear kernel for non-linear model.

62 Non-linear Latent Variable Models Exponentiated Quadratic (EQ) Covariance The EQ covariance has the form k i,j = k ( x i,:, x j,: ), where k ( ) 2 x i,: x j,: x i,:, x j,: = α exp 2 2l 2. No longer possible to optimise wrt X via an eigenvalue problem. Instead find gradients with respect to X, α, l and σ 2 and optimise using conjugate gradients.

63 Applications Style Based Inverse Kinematics Facilitating animation through modeling human motion (Grochow et al., 2004) Tracking Tracking using human motion models (Urtasun et al., 2005, 2006) Assisted Animation Generalizing drawings for animation (Baxter and Anjyo, 2006) Shape Models Inferring shape (e.g. pose from silhouette). (Ek et al., 2008b,a; Priacuriu and Reid, 2011a,b)

64 Stick Man Generalization with less Data than Dimensions Powerful uncertainly handling of GPs leads to surprising properties. Non-linear models can be used where there are fewer data points than dimensions without overfitting. Example: Modelling a stick man in 102 dimensions with 55 data points!

65 Stick Man II demstick Figure: The latent space for the stick man motion capture data.

66 Selecting Data Dimensionality GP-LVM Provides probabilistic non-linear dimensionality reduction. How to select the dimensionality? Need to estimate marginal likelihood. In standard GP-LVM it increases with increasing q.

67 Integrate Mapping Function and Latent Variables Bayesian GP-LVM Start with a standard GP-LVM. X Y σ 2 p p (Y X) = N ( y :,j 0, K ) j=1

68 Integrate Mapping Function and Latent Variables Bayesian GP-LVM Start with a standard GP-LVM. Apply standard latent variable approach: Define Gaussian prior over latent space, X. X Y σ 2 p p (Y X) = N ( y :,j 0, K ) j=1

69 Integrate Mapping Function and Latent Variables Bayesian GP-LVM Start with a standard GP-LVM. Apply standard latent variable approach: Define Gaussian prior over latent space, X. Integrate out latent variables. X Y σ 2 p p (Y X) = N ( y :,j 0, K ) j=1 p (X) = q j=1 N ( x :,j 0, α 2 i I )

70 Integrate Mapping Function and Latent Variables Bayesian GP-LVM Start with a standard GP-LVM. Apply standard latent variable approach: Define Gaussian prior over latent space, X. Integrate out latent variables. Unfortunately integration is intractable. X Y p (Y X) = p (X) = q j=1 σ 2 p N ( y :,j 0, K ) j=1 N ( x :,j 0, α 2 i I ) p (Y α) =??

71 Standard Variational Approach Fails Standard variational bound has the form: L = log p(y X) q(x) + KL ( q(x) p(x) )

72 Standard Variational Approach Fails Standard variational bound has the form: L = log p(y X) q(x) + KL ( q(x) p(x) ) Requires expectation of log p(y X) under q(x). log p(y X) = 1 ( 2 y K f,f + σ 2 I ) 1 1 y 2 log Kf,f + σ 2 I n log 2π 2

73 Standard Variational Approach Fails Standard variational bound has the form: L = log p(y X) q(x) + KL ( q(x) p(x) ) Requires expectation of log p(y X) under q(x). log p(y X) = 1 ( 2 y K f,f + σ 2 I ) 1 1 y 2 log Kf,f + σ 2 I n log 2π 2 Extremely difficult to compute because K f,f is dependent on X and appears in the inverse.

74 Variational Bayesian GP-LVM Consider collapsed variational bound, p(y) n c i i=1 N ( y f, σ 2 I ) p(u)du

75 Variational Bayesian GP-LVM Consider collapsed variational bound, p(y X) n c i i=1 N ( y f p(f u,x), σ 2 I ) p(u)du

76 Variational Bayesian GP-LVM Consider collapsed variational bound, p(y X)p(X)dX n c i N ( y f p(f u,x), σ 2 I ) p(x)dxp(u)du i=1

77 Variational Bayesian GP-LVM Consider collapsed variational bound, p(y X)p(X)dX n c i N ( y f p(f u,x), σ 2 I ) p(x)dxp(u)du i=1 Apply variational lower bound to the inner integral.

78 Variational Bayesian GP-LVM Consider collapsed variational bound, p(y X)p(X)dX n c i N ( y f p(f u,x), σ 2 I ) p(x)dxp(u)du i=1 Apply variational lower bound to the inner integral. n c i N ( y f p(f u,x), σ 2 I ) p(x)dx i=1 n log c i i=1 q(x) + log N ( y f p(f u,x), σ 2 I ) q(x) + KL ( q(x) p(x) )

79 Variational Bayesian GP-LVM Consider collapsed variational bound, p(y X)p(X)dX n c i N ( y f p(f u,x), σ 2 I ) p(x)dxp(u)du i=1 Apply variational lower bound to the inner integral. n c i N ( y f p(f u,x), σ 2 I ) p(x)dx i=1 n log c i i=1 q(x) + log N ( y f p(f u,x), σ 2 I ) q(x) + KL ( q(x) p(x) ) Which is analytically tractable for Gaussian q(x) and some covariance functions.

80 Required Expectations Need expectations under q(x) of: and log c i = 1 [ ] ki,i 2σ 2 k i,u K 1 u,uk i,u log N ( y f p(f u,y), σ 2 I ) = 1 2 log 2πσ2 1 2σ 2 ( yi K f,u K 1 u,uu ) 2 This requires the expectations Kf,u and q(x) Kf,u K 1 u,uk u,f q(x) which can be computed analytically for some covariance functions.

81 Priors for Latent Space Titsias and Lawrence (2010) Variational marginalization of X allows us to learn parameters of p(x). Standard GP-LVM where X learnt by MAP, this is not possible (see e.g. Wang et al., 2008). First example: learn the dimensionality of latent space.

82 Graphical Representations of GP-LVM X latent space Y data space

83 Graphical Representations of GP-LVM x 1 x 2 x 3 x 4 x 5 x 6 latent space y 1 y 2 y 3 y 4 y 5 y 6 y 7 y 8 data space

84 Graphical Representations of GP-LVM x 1 x 2 x 3 x 4 x 5 x 6 latent space y 1 data space

85 Graphical Representations of GP-LVM x 1 x 2 x 3 x 4 x 5 x 6 latent space σ 2 y w data space

86 Graphical Representations of GP-LVM x 1 x 2 x 3 x 4 x 5 x 6 latent space σ 2 y w α data space w N (0, αi) x N (0, I) y N ( x w, σ 2)

87 Graphical Representations of GP-LVM α x 1 x 2 x 3 x 4 x 5 x 6 latent space σ 2 y w data space w N (0, I) x N (0, αi) y N ( x w, σ 2)

88 Graphical Representations of GP-LVM α 1 α 2 α 3 α 4 α 5 α 6 x 1 x 2 x 3 x 4 x 5 x 6 latent space σ 2 y w data space w N (0, I) x i N (0, α i ) y N ( x w, σ 2)

89 Graphical Representations of GP-LVM α 1 α 2 α 3 α 4 α 5 α 6 w 1 w 2 w 3 w 4 w 5 w 6 latent space σ 2 y x data space w i N (0, α i ) x N (0, I) y N ( x w, σ 2)

90 Non-linear f (x) In linear case equivalence because f (x) = w x p(w i ) N (0, α i ) In non linear case, need to scale columns of X in prior for f (x). This implies scaling columns of X in covariance function ( k(x i,:, x j,: ) = exp 1 ) 2 (x :,i x :,j ) A(x :,i x :,j ) A is diagonal with elements α 2. Now keep prior spherical i p (X) = q N ( x :,j 0, I ) j=1 Covariance functions of this type are known as ARD (see e.g. Neal, 1996; MacKay, 2003; Rasmussen and Williams, 2006).

91 Other Priors on X Dynamical prior gives us Gaussian process dynamical system (Wang et al., 2006; Damianou et al., 2011) Structured learning prior gives us (soft) manifold sharing (Shon et al., 2006; Navaratnam et al., 2007; Ek et al., 2008b,a; Damianou et al., 2012) Gaussian process prior gives us Deep Gaussian Processes (Lawrence and Moore, 2007; Damianou and Lawrence, 2013)

92 References I W. V. Baxter and K.-I. Anjyo. Latent doodle space. In EUROGRAPHICS, volume 25, pages , Vienna, Austria, September A. Damianou, C. H. Ek, M. K. Titsias, and N. D. Lawrence. Manifold relevance determination. In J. Langford and J. Pineau, editors, Proceedings of the International Conference in Machine Learning, volume 29, San Francisco, CA, Morgan Kauffman. [PDF]. A. Damianou and N. D. Lawrence. Deep Gaussian processes. In C. Carvalho and P. Ravikumar, editors, Proceedings of the Sixteenth International Workshop on Artificial Intelligence and Statistics, volume 31, pages , AZ, USA, JMLR W&CP 31. [PDF]. A. Damianou, M. K. Titsias, and N. D. Lawrence. Variational Gaussian process dynamical systems. In P. Bartlett, F. Peirrera, C. K. I. Williams, and J. Lafferty, editors, Advances in Neural Information Processing Systems, volume 24, Cambridge, MA, MIT Press. [PDF]. C. H. Ek, J. Rihan, P. H. S. Torr, G. Rogez, and N. D. Lawrence. Ambiguity modeling in latent spaces. In A. Popescu-Belis and R. Stiefelhagen, editors, Machine Learning for Multimodal Interaction (MLMI 2008), LNCS, pages Springer-Verlag, June 2008a. [PDF]. C. H. Ek, P. H. S. Torr, and N. D. Lawrence. Gaussian process latent variable models for human pose estimation. In A. Popescu-Belis, S. Renals, and H. Bourlard, editors, Machine Learning for Multimodal Interaction (MLMI 2007), volume 4892 of LNCS, pages , Brno, Czech Republic, 2008b. Springer-Verlag. [PDF]. K. Grochow, S. L. Martin, A. Hertzmann, and Z. Popovic. Style-based inverse kinematics. In ACM Transactions on Graphics (SIGGRAPH 2004), pages , N. D. Lawrence. Gaussian process models for visualisation of high dimensional data. In S. Thrun, L. Saul, and B. Schölkopf, editors, Advances in Neural Information Processing Systems, volume 16, pages , Cambridge, MA, MIT Press. N. D. Lawrence. Probabilistic non-linear principal component analysis with Gaussian process latent variable models. Journal of Machine Learning Research, 6: , N. D. Lawrence and A. J. Moore. Hierarchical Gaussian process latent variable models. In Z. Ghahramani, editor, Proceedings of the International Conference in Machine Learning, volume 24, pages Omnipress, [Google Books]. [PDF].

93 References II D. J. C. MacKay. Information Theory, Inference and Learning Algorithms. Cambridge University Press, Cambridge, U.K., [Google Books]. R. Navaratnam, A. Fitzgibbon, and R. Cipolla. The joint manifold model for semi-supervised multi-valued regression. In IEEE International Conference on Computer Vision (ICCV). IEEE Computer Society Press, R. M. Neal. Bayesian Learning for Neural Networks. Springer, Lecture Notes in Statistics 118. V. Priacuriu and I. D. Reid. Nonlinear shape manifolds as shape priors in level set segmentation and tracking. In Proceedings of the IEEE Computer Society Conference on Computer Vision and Pattern Recognition, 2011a. V. Priacuriu and I. D. Reid. Shared shape spaces. In IEEE International Conference on Computer Vision (ICCV), 2011b. C. E. Rasmussen and C. K. I. Williams. Gaussian Processes for Machine Learning. MIT Press, Cambridge, MA, [Google Books]. A. P. Shon, K. Grochow, A. Hertzmann, and R. P. N. Rao. Learning shared latent structure for image synthesis and robotic imitation. In Weiss et al. (2006). M. E. Tipping and C. M. Bishop. Probabilistic principal component analysis. Journal of the Royal Statistical Society, B, 6 (3): , [PDF]. [DOI]. M. K. Titsias and N. D. Lawrence. Bayesian Gaussian process latent variable model. In Y. W. Teh and D. M. Titterington, editors, Proceedings of the Thirteenth International Workshop on Artificial Intelligence and Statistics, volume 9, pages , Chia Laguna Resort, Sardinia, Italy, May JMLR W&CP 9. [PDF]. R. Urtasun, D. J. Fleet, and P. Fua. 3D people tracking with Gaussian process dynamical models. In Proceedings of the IEEE Computer Society Conference on Computer Vision and Pattern Recognition, pages , New York, U.S.A., Jun IEEE Computer Society Press. R. Urtasun, D. J. Fleet, A. Hertzmann, and P. Fua. Priors for people tracking from small training sets. In IEEE International Conference on Computer Vision (ICCV), pages , Bejing, China, Oct IEEE Computer Society Press. J. M. Wang, D. J. Fleet, and A. Hertzmann. Gaussian process dynamical models. In Weiss et al. (2006). J. M. Wang, D. J. Fleet, and A. Hertzmann. Gaussian process dynamical models for human motion. IEEE Transactions on Pattern Analysis and Machine Intelligence, 30(2): , ISSN [DOI]. Y. Weiss, B. Schölkopf, and J. C. Platt, editors. Advances in Neural Information Processing Systems, volume 18, Cambridge, MA, MIT Press.