Log Gaussian Cox Processes. Chi Group Meeting February 23, 2016

Transcription

1 Log Gaussian Cox Processes Chi Group Meeting February 23, 2016

2 Outline Typical motivating application Introduction to LGCP model Brief overview of inference Applications in my work just getting started with this

3 Motivation Image: Lloyd, et al ICML.

4 Outline Typical motivating application Introduction to LGCP model Brief overview of inference Applications in my work

5 Value What are Gaussian processes? Time

6 What are Gaussian processes? Value Posterior = Likelihood * prior Marginal likelihood Time Infinite dimensional Gaussian likelihood prior

7 Value What are Gaussian processes? Time Infinite dimensional Gaussian likelihood prior

8 What is a Poisson process? Rate of point appearance: λ(t) Time

9 What is a Poisson process? Rate of point appearance: λ(t) Time Time

10 What is a Log Gaussian Cox Process? doubly stochastic Poisson process Gaussian process modulated Poisson process sigmoidal Gaussian Cox process Time

11 What is a Log Gaussian Cox Process? doubly stochastic Poisson process Cox process = Gaussian process modulated Poisson process inhomogeneous Poisson process with stochastic sigmoidal Gaussian Cox process intensity Time

12 What is a Log Gaussian Cox Process? doubly stochastic Poisson process Cox process = Gaussian process modulated Poisson process inhomogeneous Poisson process with stochastic sigmoidal Gaussian Cox process intensity Time

13 What is a Log Gaussian Cox Process? doubly stochastic Poisson process Cox process = Gaussian process modulated Poisson process inhomogeneous Poisson process with stochastic sigmoidal Gaussian Cox process intensity Always need a positive intensity, so - Take exponential - Sigmoid - Square Time

14 Outline Typical motivating application Introduction to LGCP model Brief overview of inference Applications in my work

15 How can I do inference with this sort of model? IEB Stats

16 More model specifications Number of points inside a given spacetime region: s [Flaxman, et al Fast Kronecker inference in Gaussian processes with non-gaussian likelihoods. ICML. ]

17 More model specifications Number of points inside a given spacetime region: i Simplify by introducing a spatial grid, y i = count of points inside grid cell i [Flaxman, et al Fast Kronecker inference in Gaussian processes with non-gaussian likelihoods. ICML. ]

18 How can I do inference with this sort of model? Laplace approximation Posterior distribution,

19 How can I do inference with this sort of model? Laplace approximation Posterior distribution, Gaussian approximation z 0

20 How can I do inference with this sort of model? Laplace approximation Posterior distribution, Gaussian approximation z 0 If assume a Normal centered at x 0, and take Taylor series expansion around x 0, math works out to show that Gaussian approximation of distribution is: posterior =

21 How can I do inference with this sort of model? Laplace approximation Posterior distribution, Gaussian approximation Need: - Find maximum of posterior - Hessian (-A) at maximum z 0 If assume a Normal centered at x 0, and take Taylor series expansion around x 0, math works out to show that Gaussian approximation of distribution is: posterior =

22 How can I do inference with this sort of model? Laplace approximation + Kronecker methods Can decompose GP kernel as a product of covariance matrices (because on grid) Need to do lots of inversions and logdeterminants when doing GP regression Kronecker methods can speed this up quite a bit

23 How can I do inference with this sort of model? Variational Bayes No grid required, but do need inducing points ( which can best be set on a rectangular grid) [Lloyd, et al Variational inference for Gaussian process modulated Poisson processes ICML. ] Sampling Metropolis Hastings x 2 Hamiltonian Monte Carlo x 2 [Adams, et al Tractable nonparametric Bayesian inference in Poisson processes with Gaussian process intensities. ICML. ]

24 Outline Typical motivating applications Introduction to LGCP model Brief overview of inference Applications in my work

25 References of interest Transformation Inference Notes Adams, et al Tractable nonparametric Bayesian inference in Poisson processes with Gaussian process intensities. ICML. Sigmoid function Multiple sampling schemes Flaxman, et al Fast Kronecker inference in Gaussian processes with non-gaussian likelihoods. ICML. Exponential Laplace approximation Implemented in GPML: du/~andrewgw/patte rn/ Gunter, et al Efficient Bayesian nonarametric modelling of structured point processes. UAI. Sigmoid function Many sampling schemes Multiple realizations from latent LCGP Lloyd, et al Variational inference for Gaussian process modulated Poisson processes ICML. Square Variational Bayes

26 That s all