Deep Gaussian Processes


at MLSS, Stellenbosch, South Africa on Jan 11, 2019 [jupyter][reveal]
Neil D. Lawrence, Amazon Cambridge and University of Sheffield


Gaussian process models provide a flexible, non-parametric approach to modelling that sustains uncertainty about the function. However, computational demands and the joint Gaussian assumption make them inappropriate for some applications. In this talk we review low rank approximations for Gaussian processes and use stochastic process composition to create non-Gaussian processes. We illustrate the models on simple regression tasks to give a sense of how uncertainty propagates through the model. We end will demonstrations on unsupervised learning of digits and motion capture data.

$$\newcommand{\Amatrix}{\mathbf{A}} \newcommand{\KL}[2]{\text{KL}\left( #1\,\|\,#2 \right)} \newcommand{\Kaast}{\kernelMatrix_{\mathbf{ \ast}\mathbf{ \ast}}} \newcommand{\Kastu}{\kernelMatrix_{\mathbf{ \ast} \inducingVector}} \newcommand{\Kff}{\kernelMatrix_{\mappingFunctionVector \mappingFunctionVector}} \newcommand{\Kfu}{\kernelMatrix_{\mappingFunctionVector \inducingVector}} \newcommand{\Kuast}{\kernelMatrix_{\inducingVector \bf\ast}} \newcommand{\Kuf}{\kernelMatrix_{\inducingVector \mappingFunctionVector}} \newcommand{\Kuu}{\kernelMatrix_{\inducingVector \inducingVector}} \newcommand{\Kuui}{\Kuu^{-1}} \newcommand{\Qaast}{\mathbf{Q}_{\bf \ast \ast}} \newcommand{\Qastf}{\mathbf{Q}_{\ast \mappingFunction}} \newcommand{\Qfast}{\mathbf{Q}_{\mappingFunctionVector \bf \ast}} \newcommand{\Qff}{\mathbf{Q}_{\mappingFunctionVector \mappingFunctionVector}} \newcommand{\aMatrix}{\mathbf{A}} \newcommand{\aScalar}{a} \newcommand{\aVector}{\mathbf{a}} \newcommand{\acceleration}{a} \newcommand{\bMatrix}{\mathbf{B}} \newcommand{\bScalar}{b} \newcommand{\bVector}{\mathbf{b}} \newcommand{\basisFunc}{\phi} \newcommand{\basisFuncVector}{\boldsymbol{ \basisFunc}} \newcommand{\basisFunction}{\phi} \newcommand{\basisLocation}{\mu} \newcommand{\basisMatrix}{\boldsymbol{ \Phi}} \newcommand{\basisScalar}{\basisFunction} \newcommand{\basisVector}{\boldsymbol{ \basisFunction}} \newcommand{\activationFunction}{\phi} \newcommand{\activationMatrix}{\boldsymbol{ \Phi}} \newcommand{\activationScalar}{\basisFunction} \newcommand{\activationVector}{\boldsymbol{ \basisFunction}} \newcommand{\bigO}{\mathcal{O}} \newcommand{\binomProb}{\pi} \newcommand{\cMatrix}{\mathbf{C}} \newcommand{\cbasisMatrix}{\hat{\boldsymbol{ \Phi}}} \newcommand{\cdataMatrix}{\hat{\dataMatrix}} \newcommand{\cdataScalar}{\hat{\dataScalar}} \newcommand{\cdataVector}{\hat{\dataVector}} \newcommand{\centeredKernelMatrix}{\mathbf{ \MakeUppercase{\centeredKernelScalar}}} \newcommand{\centeredKernelScalar}{b} \newcommand{\centeredKernelVector}{\centeredKernelScalar} \newcommand{\centeringMatrix}{\mathbf{H}} \newcommand{\chiSquaredDist}[2]{\chi_{#1}^{2}\left(#2\right)} \newcommand{\chiSquaredSamp}[1]{\chi_{#1}^{2}} \newcommand{\conditionalCovariance}{\boldsymbol{ \Sigma}} \newcommand{\coregionalizationMatrix}{\mathbf{B}} \newcommand{\coregionalizationScalar}{b} \newcommand{\coregionalizationVector}{\mathbf{ \coregionalizationScalar}} \newcommand{\covDist}[2]{\text{cov}_{#2}\left(#1\right)} \newcommand{\covSamp}[1]{\text{cov}\left(#1\right)} \newcommand{\covarianceScalar}{c} \newcommand{\covarianceVector}{\mathbf{ \covarianceScalar}} \newcommand{\covarianceMatrix}{\mathbf{C}} \newcommand{\covarianceMatrixTwo}{\boldsymbol{ \Sigma}} \newcommand{\croupierScalar}{s} \newcommand{\croupierVector}{\mathbf{ \croupierScalar}} \newcommand{\croupierMatrix}{\mathbf{ \MakeUppercase{\croupierScalar}}} \newcommand{\dataDim}{p} \newcommand{\dataIndex}{i} \newcommand{\dataIndexTwo}{j} \newcommand{\dataMatrix}{\mathbf{Y}} \newcommand{\dataScalar}{y} \newcommand{\dataSet}{\mathcal{D}} \newcommand{\dataStd}{\sigma} \newcommand{\dataVector}{\mathbf{ \dataScalar}} \newcommand{\decayRate}{d} \newcommand{\degreeMatrix}{\mathbf{ \MakeUppercase{\degreeScalar}}} \newcommand{\degreeScalar}{d} \newcommand{\degreeVector}{\mathbf{ \degreeScalar}} % Already defined by latex %\newcommand{\det}[1]{\left|#1\right|} \newcommand{\diag}[1]{\text{diag}\left(#1\right)} \newcommand{\diagonalMatrix}{\mathbf{D}} \newcommand{\diff}[2]{\frac{\text{d}#1}{\text{d}#2}} \newcommand{\diffTwo}[2]{\frac{\text{d}^2#1}{\text{d}#2^2}} \newcommand{\displacement}{x} \newcommand{\displacementVector}{\textbf{\displacement}} \newcommand{\distanceMatrix}{\mathbf{ \MakeUppercase{\distanceScalar}}} \newcommand{\distanceScalar}{d} \newcommand{\distanceVector}{\mathbf{ \distanceScalar}} \newcommand{\eigenvaltwo}{\ell} \newcommand{\eigenvaltwoMatrix}{\mathbf{L}} \newcommand{\eigenvaltwoVector}{\mathbf{l}} \newcommand{\eigenvalue}{\lambda} \newcommand{\eigenvalueMatrix}{\boldsymbol{ \Lambda}} \newcommand{\eigenvalueVector}{\boldsymbol{ \lambda}} \newcommand{\eigenvector}{\mathbf{ \eigenvectorScalar}} \newcommand{\eigenvectorMatrix}{\mathbf{U}} \newcommand{\eigenvectorScalar}{u} \newcommand{\eigenvectwo}{\mathbf{v}} \newcommand{\eigenvectwoMatrix}{\mathbf{V}} \newcommand{\eigenvectwoScalar}{v} \newcommand{\entropy}[1]{\mathcal{H}\left(#1\right)} \newcommand{\errorFunction}{E} \newcommand{\expDist}[2]{\left<#1\right>_{#2}} \newcommand{\expSamp}[1]{\left<#1\right>} \newcommand{\expectation}[1]{\left\langle #1 \right\rangle } \newcommand{\expectationDist}[2]{\left\langle #1 \right\rangle _{#2}} \newcommand{\expectedDistanceMatrix}{\mathcal{D}} \newcommand{\eye}{\mathbf{I}} \newcommand{\fantasyDim}{r} \newcommand{\fantasyMatrix}{\mathbf{ \MakeUppercase{\fantasyScalar}}} \newcommand{\fantasyScalar}{z} \newcommand{\fantasyVector}{\mathbf{ \fantasyScalar}} \newcommand{\featureStd}{\varsigma} \newcommand{\gammaCdf}[3]{\mathcal{GAMMA CDF}\left(#1|#2,#3\right)} \newcommand{\gammaDist}[3]{\mathcal{G}\left(#1|#2,#3\right)} \newcommand{\gammaSamp}[2]{\mathcal{G}\left(#1,#2\right)} \newcommand{\gaussianDist}[3]{\mathcal{N}\left(#1|#2,#3\right)} \newcommand{\gaussianSamp}[2]{\mathcal{N}\left(#1,#2\right)} \newcommand{\given}{|} \newcommand{\half}{\frac{1}{2}} \newcommand{\heaviside}{H} \newcommand{\hiddenMatrix}{\mathbf{ \MakeUppercase{\hiddenScalar}}} \newcommand{\hiddenScalar}{h} \newcommand{\hiddenVector}{\mathbf{ \hiddenScalar}} \newcommand{\identityMatrix}{\eye} \newcommand{\inducingInputScalar}{z} \newcommand{\inducingInputVector}{\mathbf{ \inducingInputScalar}} \newcommand{\inducingInputMatrix}{\mathbf{Z}} \newcommand{\inducingScalar}{u} \newcommand{\inducingVector}{\mathbf{ \inducingScalar}} \newcommand{\inducingMatrix}{\mathbf{U}} \newcommand{\inlineDiff}[2]{\text{d}#1/\text{d}#2} \newcommand{\inputDim}{q} \newcommand{\inputMatrix}{\mathbf{X}} \newcommand{\inputScalar}{x} \newcommand{\inputSpace}{\mathcal{X}} \newcommand{\inputVals}{\inputVector} \newcommand{\inputVector}{\mathbf{ \inputScalar}} \newcommand{\iterNum}{k} \newcommand{\kernel}{\kernelScalar} \newcommand{\kernelMatrix}{\mathbf{K}} \newcommand{\kernelScalar}{k} \newcommand{\kernelVector}{\mathbf{ \kernelScalar}} \newcommand{\kff}{\kernelScalar_{\mappingFunction \mappingFunction}} \newcommand{\kfu}{\kernelVector_{\mappingFunction \inducingScalar}} \newcommand{\kuf}{\kernelVector_{\inducingScalar \mappingFunction}} \newcommand{\kuu}{\kernelVector_{\inducingScalar \inducingScalar}} \newcommand{\lagrangeMultiplier}{\lambda} \newcommand{\lagrangeMultiplierMatrix}{\boldsymbol{ \Lambda}} \newcommand{\lagrangian}{L} \newcommand{\laplacianFactor}{\mathbf{ \MakeUppercase{\laplacianFactorScalar}}} \newcommand{\laplacianFactorScalar}{m} \newcommand{\laplacianFactorVector}{\mathbf{ \laplacianFactorScalar}} \newcommand{\laplacianMatrix}{\mathbf{L}} \newcommand{\laplacianScalar}{\ell} \newcommand{\laplacianVector}{\mathbf{ \ell}} \newcommand{\latentDim}{q} \newcommand{\latentDistanceMatrix}{\boldsymbol{ \Delta}} \newcommand{\latentDistanceScalar}{\delta} \newcommand{\latentDistanceVector}{\boldsymbol{ \delta}} \newcommand{\latentForce}{f} \newcommand{\latentFunction}{u} \newcommand{\latentFunctionVector}{\mathbf{ \latentFunction}} \newcommand{\latentFunctionMatrix}{\mathbf{ \MakeUppercase{\latentFunction}}} \newcommand{\latentIndex}{j} \newcommand{\latentScalar}{z} \newcommand{\latentVector}{\mathbf{ \latentScalar}} \newcommand{\latentMatrix}{\mathbf{Z}} \newcommand{\learnRate}{\eta} \newcommand{\lengthScale}{\ell} \newcommand{\rbfWidth}{\ell} \newcommand{\likelihoodBound}{\mathcal{L}} \newcommand{\likelihoodFunction}{L} \newcommand{\locationScalar}{\mu} \newcommand{\locationVector}{\boldsymbol{ \locationScalar}} \newcommand{\locationMatrix}{\mathbf{M}} \newcommand{\variance}[1]{\text{var}\left( #1 \right)} \newcommand{\mappingFunction}{f} \newcommand{\mappingFunctionMatrix}{\mathbf{F}} \newcommand{\mappingFunctionTwo}{g} \newcommand{\mappingFunctionTwoMatrix}{\mathbf{G}} \newcommand{\mappingFunctionTwoVector}{\mathbf{ \mappingFunctionTwo}} \newcommand{\mappingFunctionVector}{\mathbf{ \mappingFunction}} \newcommand{\scaleScalar}{s} \newcommand{\mappingScalar}{w} \newcommand{\mappingVector}{\mathbf{ \mappingScalar}} \newcommand{\mappingMatrix}{\mathbf{W}} \newcommand{\mappingScalarTwo}{v} \newcommand{\mappingVectorTwo}{\mathbf{ \mappingScalarTwo}} \newcommand{\mappingMatrixTwo}{\mathbf{V}} \newcommand{\maxIters}{K} \newcommand{\meanMatrix}{\mathbf{M}} \newcommand{\meanScalar}{\mu} \newcommand{\meanTwoMatrix}{\mathbf{M}} \newcommand{\meanTwoScalar}{m} \newcommand{\meanTwoVector}{\mathbf{ \meanTwoScalar}} \newcommand{\meanVector}{\boldsymbol{ \meanScalar}} \newcommand{\mrnaConcentration}{m} \newcommand{\naturalFrequency}{\omega} \newcommand{\neighborhood}[1]{\mathcal{N}\left( #1 \right)} \newcommand{\neilurl}{} \newcommand{\noiseMatrix}{\boldsymbol{ E}} \newcommand{\noiseScalar}{\epsilon} \newcommand{\noiseVector}{\boldsymbol{ \epsilon}} \newcommand{\norm}[1]{\left\Vert #1 \right\Vert} \newcommand{\normalizedLaplacianMatrix}{\hat{\mathbf{L}}} \newcommand{\normalizedLaplacianScalar}{\hat{\ell}} \newcommand{\normalizedLaplacianVector}{\hat{\mathbf{ \ell}}} \newcommand{\numActive}{m} \newcommand{\numBasisFunc}{m} \newcommand{\numComponents}{m} \newcommand{\numComps}{K} \newcommand{\numData}{n} \newcommand{\numFeatures}{K} \newcommand{\numHidden}{h} \newcommand{\numInducing}{m} \newcommand{\numLayers}{\ell} \newcommand{\numNeighbors}{K} \newcommand{\numSequences}{s} \newcommand{\numSuccess}{s} \newcommand{\numTasks}{m} \newcommand{\numTime}{T} \newcommand{\numTrials}{S} \newcommand{\outputIndex}{j} \newcommand{\paramVector}{\boldsymbol{ \theta}} \newcommand{\parameterMatrix}{\boldsymbol{ \Theta}} \newcommand{\parameterScalar}{\theta} \newcommand{\parameterVector}{\boldsymbol{ \parameterScalar}} \newcommand{\partDiff}[2]{\frac{\partial#1}{\partial#2}} \newcommand{\precisionScalar}{j} \newcommand{\precisionVector}{\mathbf{ \precisionScalar}} \newcommand{\precisionMatrix}{\mathbf{J}} \newcommand{\pseudotargetScalar}{\widetilde{y}} \newcommand{\pseudotargetVector}{\mathbf{ \pseudotargetScalar}} \newcommand{\pseudotargetMatrix}{\mathbf{ \widetilde{Y}}} \newcommand{\rank}[1]{\text{rank}\left(#1\right)} \newcommand{\rayleighDist}[2]{\mathcal{R}\left(#1|#2\right)} \newcommand{\rayleighSamp}[1]{\mathcal{R}\left(#1\right)} \newcommand{\responsibility}{r} \newcommand{\rotationScalar}{r} \newcommand{\rotationVector}{\mathbf{ \rotationScalar}} \newcommand{\rotationMatrix}{\mathbf{R}} \newcommand{\sampleCovScalar}{s} \newcommand{\sampleCovVector}{\mathbf{ \sampleCovScalar}} \newcommand{\sampleCovMatrix}{\mathbf{s}} \newcommand{\scalarProduct}[2]{\left\langle{#1},{#2}\right\rangle} \newcommand{\sign}[1]{\text{sign}\left(#1\right)} \newcommand{\sigmoid}[1]{\sigma\left(#1\right)} \newcommand{\singularvalue}{\ell} \newcommand{\singularvalueMatrix}{\mathbf{L}} \newcommand{\singularvalueVector}{\mathbf{l}} \newcommand{\sorth}{\mathbf{u}} \newcommand{\spar}{\lambda} \newcommand{\trace}[1]{\text{tr}\left(#1\right)} \newcommand{\BasalRate}{B} \newcommand{\DampingCoefficient}{C} \newcommand{\DecayRate}{D} \newcommand{\Displacement}{X} \newcommand{\LatentForce}{F} \newcommand{\Mass}{M} \newcommand{\Sensitivity}{S} \newcommand{\basalRate}{b} \newcommand{\dampingCoefficient}{c} \newcommand{\mass}{m} \newcommand{\sensitivity}{s} \newcommand{\springScalar}{\kappa} \newcommand{\springVector}{\boldsymbol{ \kappa}} \newcommand{\springMatrix}{\boldsymbol{ \mathcal{K}}} \newcommand{\tfConcentration}{p} \newcommand{\tfDecayRate}{\delta} \newcommand{\tfMrnaConcentration}{f} \newcommand{\tfVector}{\mathbf{ \tfConcentration}} \newcommand{\velocity}{v} \newcommand{\sufficientStatsScalar}{g} \newcommand{\sufficientStatsVector}{\mathbf{ \sufficientStatsScalar}} \newcommand{\sufficientStatsMatrix}{\mathbf{G}} \newcommand{\switchScalar}{s} \newcommand{\switchVector}{\mathbf{ \switchScalar}} \newcommand{\switchMatrix}{\mathbf{S}} \newcommand{\tr}[1]{\text{tr}\left(#1\right)} \newcommand{\loneNorm}[1]{\left\Vert #1 \right\Vert_1} \newcommand{\ltwoNorm}[1]{\left\Vert #1 \right\Vert_2} \newcommand{\onenorm}[1]{\left\vert#1\right\vert_1} \newcommand{\twonorm}[1]{\left\Vert #1 \right\Vert} \newcommand{\vScalar}{v} \newcommand{\vVector}{\mathbf{v}} \newcommand{\vMatrix}{\mathbf{V}} \newcommand{\varianceDist}[2]{\text{var}_{#2}\left( #1 \right)} % Already defined by latex %\newcommand{\vec}{#1:} \newcommand{\vecb}[1]{\left(#1\right):} \newcommand{\weightScalar}{w} \newcommand{\weightVector}{\mathbf{ \weightScalar}} \newcommand{\weightMatrix}{\mathbf{W}} \newcommand{\weightedAdjacencyMatrix}{\mathbf{A}} \newcommand{\weightedAdjacencyScalar}{a} \newcommand{\weightedAdjacencyVector}{\mathbf{ \weightedAdjacencyScalar}} \newcommand{\onesVector}{\mathbf{1}} \newcommand{\zerosVector}{\mathbf{0}} $$



In the previous session on Gaussian processes, we introduced the Gaussian process model and the covariance function. In this session we are going to address two challenges of the Gaussian process. Firstly, we look at the computational tractability and secondly we look at extending the nature of the process beyond Gaussian.

Low Rank Gaussian Processes

Image credit: Kai Arulkumaran

Inference in a Gaussian process has computational complexity of $\bigO(\numData^3)$ and storage demands of $\bigO(\numData^2)$. This is too large for many modern data sets.

Low rank approximations allow us to work with Gaussian processes with computational complexity of $\bigO(\numData\numInducing^2)$ and storage demands of $\bigO(\numData\numInducing)$, where $\numInducing$ is a user chosen parameter.

In machine learning, low rank approximations date back to Smola and Bartlett (n.d.), Williams and Seeger (n.d.), who considered the Nystr"om approximation and Csató and Opper (2002);Csató (2002) who considered low rank approximations in the context of on-line learning. Selection of active points for the approximation was considered by Seeger, Williams, and Lawrence (n.d.) and Snelson and Ghahramani (n.d.) first proposed that the active set could be optimized directly. Those approaches were reviewed by Quiñonero Candela and Rasmussen (2005) under a unifying likelihood approximation perspective. General rules for deriving the maximum likelihood for these sparse approximations were given in Lawrence (n.d.).

Modern variational interpretations of these low rank approaches were first explored in Titsias (n.d.). A more modern summary which considers each of these approximations as an α-divergence is given by Bui, Yan, and Turner (2017).

## Variational Compression

Inducing variables are a compression of the real observations. The basic idea is can I create a new data set that summarizes all the information in the original data set. If this data set is smaller, I've compressed the information in the original data set.

Inducing variables can be thought of as pseudo-data, indeed in Snelson and Ghahramani (n.d.) they were referred to as pseudo-points.

The only requirement for inducing variables is that they are jointly distributed as a Gaussian process with the original data. This means that they can be from the space $\mappingFunctionVector$ or a space that is related through a linear operator (see e.g. Álvarez et al. (2010)). For example we could choose to store the gradient of the function at particular points or a value from the frequency spectrum of the function (Lázaro-Gredilla, Quiñonero-Candela, and Rasmussen 2010).

Variational Compression II

Inducing variables don't only allow for the compression of the non-parameteric information into a reduced data aset but they also allow for computational scaling of the algorithms through, for example stochastic variational approaches Hensman, Fusi, and Lawrence (n.d.) or parallelization Gal, Wilk, and Rasmussen (n.d.),Dai et al. (2014), M. W. Seeger et al. (2017).

We’ve seen how we go from parametric to non-parametric. The limit implies infinite dimensional $\mappingVector$. Gaussian processes are generally non-parametric: combine data with covariance function to get model. This representation cannot be summarized by a parameter vector of a fixed size.

Parametric models have a representation that does not respond to increasing training set size. Bayesian posterior distributions over parameters contain the information about the training data, for example if we use use Bayes’ rule from training data,
$$ p\left(\mappingVector|\dataVector, \inputMatrix\right), $$
to make predictions on test data
$$ p\left(\dataScalar_*|\inputMatrix_*, \dataVector, \inputMatrix\right) = \int p\left(\dataScalar_*|\mappingVector,\inputMatrix_*\right)p\left(\mappingVector|\dataVector, \inputMatrix)\text{d}\mappingVector\right) $$
then $\mappingVector$ becomes a bottleneck for information about the training set to pass to the test set. The solution is to increase $\numBasisFunc$ so that the bottleneck is so large that it no longer presents a problem. How big is big enough for $\numBasisFunc$? Non-parametrics says $\numBasisFunc \rightarrow \infty$.

Now no longer possible to manipulate the model through the standard parametric form. However, it is possible to express parametric as GPs:
$$ \kernelScalar\left(\inputVector_i,\inputVector_j\right)=\basisFunction_:\left(\inputVector_i\right)^\top\basisFunction_:\left(\inputVector_j\right). $$
These are known as degenerate covariance matrices. Their rank is at most $\numBasisFunc$, non-parametric models have full rank covariance matrices. Most well known is the “linear kernel”,
$$ \kernelScalar(\inputVector_i, \inputVector_j) = \inputVector_i^\top\inputVector_j. $$
For non-parametrics prediction at a new point, $\mappingFunctionVector_*$, is made by conditioning on $\mappingFunctionVector$ in the joint distribution. In GPs this involves combining the training data with the covariance function and the mean function. Parametric is a special case when conditional prediction can be summarized in a fixed number of parameters. Complexity of parametric model remains fixed regardless of the size of our training data set. For a non-parametric model the required number of parameters grows with the size of the training data.

Augment Variable Space

In inducing variable approximations, we augment the variable space with a set of inducing points, $\inducingVector$. These inducing points are jointly Gaussian distributed with the points from our function, $\mappingFunctionVector$. So we have a joint Gaussian process with covariance,
$$ \begin{bmatrix} \mappingFunctionVector\\ \inducingVector \end{bmatrix} \sim \gaussianSamp{\zerosVector}{\kernelMatrix} $$
where the kernel matrix itself can be decomposed into
$$ \kernelMatrix = \begin{bmatrix} \Kff & \Kfu \\ \Kuf & \Kuu \end{bmatrix} $$

This defines a joint density between the original function points, $\mappingFunctionVector$ and our inducing points, $\inducingVector$. This can be decomposed through the product rule to give.
$$ p(\mappingFunctionVector, \inducingVector) = p(\mappingFunctionVector| \inducingVector) p(\inducingVector) $$
The Gaussian process is (typically) given by a noise corrupted form of $\mappingFunctionVector$, i.e.,
$$ \dataScalar(\inputVector) = \mappingFunction(\inputVector) + \noiseScalar, $$
which can be written probabilisticlly as,
$$ p(\dataVector) = \int p(\dataVector|\mappingFunctionVector) p(\mappingFunctionVector) \text{d}\mappingFunctionVector, $$
where for the independent case we have $p(\dataVector | \mappingFunctionVector) = \prod_{i=1}^\numData p(\dataScalar_i|\mappingFunction_i)$.

Inducing variables are like auxilliary variables in Monte Carlo algorithms. We introduce the inducing variables by augmenting this integral with an additional integral over $\inducingVector$,
$$ p(\dataVector) = \int p(\dataVector|\mappingFunctionVector) p(\mappingFunctionVector|\inducingVector) p(\inducingVector) \text{d}\inducingVector \text{d}\mappingFunctionVector. $$
Now, conceptually speaking we are going to integrate out $\mappingFunctionVector$, initially leaving $\inducingVector$ in place. This gives,
$$ p(\dataVector) = \int p(\dataVector|\inducingVector) p(\inducingVector) \text{d}\inducingVector. $$

Note the similarity between this form and our standard parametric form. If we had defined our model through standard basis functions we would have,
$$ \dataScalar(\inputVector) = \weightVector^\top\basisVector(\inputVector) + \noiseScalar $$
and the resulting probabilistic representation would be
$$ p(\dataVector) = \int p(\dataVector|\weightVector) p(\weightVector) \text{d} \weightVector $$
allowing us to predict
$$ p(\dataVector^*|\dataVector) = \int p(\dataVector^*|\weightVector) p(\weightVector|\dataVector) \text{d} \weightVector $$

The new prediction algorithm involves
$$ p(\dataVector^*|\dataVector) = \int p(\dataVector^*|\inducingVector) p(\inducingVector|\dataVector) \text{d} \inducingVector $$
but importantly the length of $\inducingVector$ is not fixed at design time like the number of parameters were. We can vary the number of inducing variables we use to give us the model capacity we require.

Unfortunately, computation of $p(\dataVector|\inducingVector)$ turns out to be intractable. As a result, we need to turn to approximations to make progress.

Variational Bound on $p(\dataVector |\inducingVector)$

The conditional density of the data given the inducing points can be lower bounded variationally
$$ \begin{aligned} \log p(\dataVector|\inducingVector) & = \log \int p(\dataVector|\mappingFunctionVector) p(\mappingFunctionVector|\inducingVector) \text{d}\mappingFunctionVector\\ & = \int q(\mappingFunctionVector) \log \frac{p(\dataVector|\mappingFunctionVector) p(\mappingFunctionVector|\inducingVector)}{q(\mappingFunctionVector)}\text{d}\mappingFunctionVector + \KL{q(\mappingFunctionVector)}{p(\mappingFunctionVector|\dataVector, \inducingVector)}. \end{aligned} $$

The key innovation from Titsias (n.d.) was to then make a particular choice for $q(\mappingFunctionVector)$. If we set $q(\mappingFunctionVector)=p(\mappingFunctionVector|\inducingVector)$,
$$ \log p(\dataVector|\inducingVector) \geq \int p(\mappingFunctionVector|\inducingVector) \log p(\dataVector|\mappingFunctionVector)\text{d}\mappingFunctionVector. $$

$$ p(\dataVector|\inducingVector) \geq \exp \int p(\mappingFunctionVector|\inducingVector) \log p(\dataVector|\mappingFunctionVector)\text{d}\mappingFunctionVector. $$

Optimal Compression in Inducing Variables

Maximizing the lower bound minimizes the Kullback-Leibler divergence (or information gain) between our approximating density, $p(\mappingFunctionVector|\inducingVector)$ and the true posterior density, $p(\mappingFunctionVector|\dataVector, \inducingVector)$.

$$ \KL{p(\mappingFunctionVector|\inducingVector)}{p(\mappingFunctionVector|\dataVector, \inducingVector)} = \int p(\mappingFunctionVector|\inducingVector) \log \frac{p(\mappingFunctionVector|\inducingVector)}{p(\mappingFunctionVector|\dataVector, \inducingVector)}\text{d}\inducingVector $$

This bound is minimized when the information stored about $\dataVector$ is already stored in $\inducingVector$. In other words, maximizing the bound seeks an optimal compression from the information gain perspective.

For the case where $\inducingVector = \mappingFunctionVector$ the bound is exact ($\mappingFunctionVector$ d-separates $\dataVector$ from $\inducingVector$).

Choice of Inducing Variables

The quality of the resulting bound is determined by the choice of the inducing variables. You are free to choose whichever heuristics you like for the inducing variables, as long as they are drawn jointly from a valid Gaussian process, i.e. such that
$$ \begin{bmatrix} \mappingFunctionVector\\ \inducingVector \end{bmatrix} \sim \gaussianSamp{\zerosVector}{\kernelMatrix} $$
where the kernel matrix itself can be decomposed into
$$ \kernelMatrix = \begin{bmatrix} \Kff & \Kfu \\ \Kuf & \Kuu \end{bmatrix} $$
Choosing the inducing variables amounts to specifying $\Kfu$ and $\Kuu$ such that $\kernelMatrix$ remains positive definite. The typical choice is to choose $\inducingVector$ in the same domain as $\mappingFunctionVector$, associating each inducing output, $\inducingScalar_i$ with a corresponding input location $\inducingInputVector$. However, more imaginative choices are absolutely possible. In particular, if $\inducingVector$ is related to $\mappingFunctionVector$ through a linear operator (see e.g. Álvarez et al. (2010)), then valid $\Kuu$ and $\Kuf$ can be constructed. For example we could choose to store the gradient of the function at particular points or a value from the frequency spectrum of the function (Lázaro-Gredilla, Quiñonero-Candela, and Rasmussen 2010).

Variational Compression II

Inducing variables don't only allow for the compression of the non-parameteric information into a reduced data set but they also allow for computational scaling of the algorithms through, for example stochastic variational approaches(Hoffman et al. 2012; Hensman, Fusi, and Lawrence, n.d.) or parallelization (???),Dai et al. (2014), M. W. Seeger et al. (2017).

A Simple Regression Problem

Here we set up a simple one dimensional regression problem. The input locations, $\inputMatrix$, are in two separate clusters. The response variable, $\dataVector$, is sampled from a Gaussian process with an exponentiated quadratic covariance.

import numpy as np
import GPy
N = 50
noise_var = 0.01
X = np.zeros((50, 1))
X[:25, :] = np.linspace(0,3,25)[:,None] # First cluster of inputs/covariates
X[25:, :] = np.linspace(7,10,25)[:,None] # Second cluster of inputs/covariates

# Sample response variables from a Gaussian process with exponentiated quadratic covariance.
k = GPy.kern.RBF(1)
y = np.random.multivariate_normal(np.zeros(N),k.K(X)+np.eye(N)*np.sqrt(noise_var)).reshape(-1,1)

First we perform a full Gaussian process regression on the data. We create a GP model, m_full, and fit it to the data, plotting the resulting fit.

m_full = GPy.models.GPRegression(X,y)
_ = m_full.optimize(messages=True) # Optimize parameters of covariance function
Full Gaussian process fitted to the data set.

Now we set up the inducing variables, u. Each inducing variable has its own associated input index, Z, which lives in the same space as $\inputMatrix$. Here we are using the true covariance function parameters to generate the fit.

kern = GPy.kern.RBF(1)
Z = np.hstack(
m = GPy.models.SparseGPRegression(X,y,kernel=kern,Z=Z)
m.noise_var = noise_var
Sparse Gaussian process fitted with six inducing variables, no optimization of parameters or inducing variables.
_ = m.optimize(messages=True)
Gaussian process fitted with inducing variables fixed and parameters optimized
_ = m.optimize(messages=True)
Gaussian process fitted with location of inducing variables and parameters both optimized

Now we will vary the number of inducing points used to form the approximation.

M = 8

_ = m.optimize(messages=True)
Comparison of the full Gaussian process fit with a sparse Gaussian process using eight inducing varibles. Both inducing variables and parameters are optimized.

And we can compare the probability of the result to the full model.

print(m.log_likelihood(), m_full.log_likelihood())

  • Let’s be explicity about storing approximate posterior of $\inducingVector$, $q(\inducingVector)$.
  • Now we have
    $$p(\dataVector^*|\dataVector) = \int p(\dataVector^*| \inducingVector) q(\inducingVector | \dataVector) \text{d} \inducingVector$$

  • Inducing variables look a lot like regular parameters.
  • But: their dimensionality does not need to be set at design time.
  • They can be modified arbitrarily at run time without effecting the model likelihood.
  • They only effect the quality of compression and the lower bound.

  • Exploit the resulting factorization ...
    $$p(\dataVector^*|\dataVector) = \int p(\dataVector^*| \inducingVector) q(\inducingVector | \dataVector) \text{d} \inducingVector$$
  • The distribution now factorizes:
    $$p(\dataVector^*|\dataVector) = \int \prod_{i=1}^{\numData^*}p(\dataScalar^*_i| \inducingVector) q(\inducingVector | \dataVector) \text{d} \inducingVector$$
  • This factorization can be exploited for stochastic variational inference (Hoffman et al. 2012).

Modern data availability
Proxy for index of deprivation?
Actually index of deprivation is a proxy for this ...
(Hensman, Fusi, and Lawrence, n.d.)
(Hensman, Fusi, and Lawrence, n.d.)
  • A Unifying Framework for Gaussian Process Pseudo-Point Approximations using Power Expectation Propagation Bui, Yan, and Turner (2017)

  • Deep Gaussian Processes and Variational Propagation of Uncertainty Damianou (2015)

Even in the early days of Gaussian processes in machine learning, it was understood that we were throwing something fundamental away. This is perhaps captured best by David MacKay in his 1997 NeurIPS tutorial on Gaussian processes, where he asked "Have we thrown out the baby with the bathwater?". The quote below is from his summarization paper.

According to the hype of 1987, neural networks were meant to be intelligent models which discovered features and patterns in data. Gaussian processes in contrast are simply smoothing devices. How can Gaussian processes possibly repalce neural networks? What is going on?

MacKay (n.d.)

import teaching_plots as plot
A deep neural network. Input nodes are shown at the bottom. Each hidden layer is the result of applying an affine transformation to the previous layer and placing through an activation function.

Mathematically, each layer of a neural network is given through computing the activation function, $\basisFunction(\cdot)$, contingent on the previous layer, or the inputs. In this way the activation functions, are composed to generate more complex interactions than would be possible with any single layer.
$$ \begin{align} \hiddenVector_{1} &= \basisFunction\left(\mappingMatrix_1 \inputVector\right)\\ \hiddenVector_{2} &= \basisFunction\left(\mappingMatrix_2\hiddenVector_{1}\right)\\ \hiddenVector_{3} &= \basisFunction\left(\mappingMatrix_3 \hiddenVector_{2}\right)\\ \dataVector &= \mappingVector_4 ^\top\hiddenVector_{3} \end{align} $$

## Overfitting

One potential problem is that as the number of nodes in two adjacent layers increases, the number of parameters in the affine transformation between layers, $\mappingMatrix$, increases. If there are ki − 1 nodes in one layer, and ki nodes in the following, then that matrix contains kiki − 1 parameters, when we have layer widths in the 1000s that leads to millions of parameters.

One proposed solution is known as dropout where only a sub-set of the neural network is trained at each iteration. An alternative solution would be to reparameterize $\mappingMatrix$ with its singular value decomposition.
$$ \mappingMatrix = \eigenvectorMatrix\eigenvalueMatrix\eigenvectwoMatrix^\top $$
$$ \mappingMatrix = \eigenvectorMatrix\eigenvectwoMatrix^\top $$
where if $\mappingMatrix \in \Re^{k_1\times k_2}$ then $\eigenvectorMatrix\in \Re^{k_1\times q}$ and $\eigenvectwoMatrix \in \Re^{k_2\times q}$, i.e. we have a low rank matrix factorization for the weights.

import teaching_plots as plot
Pictorial representation of the low rank form of the matrix $\mappingMatrix$

import teaching_plots as plot

Including the low rank decomposition of $\mappingMatrix$ in the neural network, we obtain a new mathematical form. Effectively, we are adding additional latent layers, $\latentVector$, in between each of the existing hidden layers. In a neural network these are sometimes known as bottleneck layers. The network can now be written mathematically as
$$ \begin{align} \latentVector_{1} &= \eigenvectwoMatrix^\top_1 \inputVector\\ \hiddenVector_{1} &= \basisFunction\left(\eigenvectorMatrix_1 \latentVector_{1}\right)\\ \latentVector_{2} &= \eigenvectwoMatrix^\top_2 \hiddenVector_{1}\\ \hiddenVector_{2} &= \basisFunction\left(\eigenvectorMatrix_2 \latentVector_{2}\right)\\ \latentVector_{3} &= \eigenvectwoMatrix^\top_3 \hiddenVector_{2}\\ \hiddenVector_{3} &= \basisFunction\left(\eigenvectorMatrix_3 \latentVector_{3}\right)\\ \dataVector &= \mappingVector_4^\top\hiddenVector_{3}. \end{align} $$

$$ \begin{align} \latentVector_{1} &= \eigenvectwoMatrix^\top_1 \inputVector\\ \latentVector_{2} &= \eigenvectwoMatrix^\top_2 \basisFunction\left(\eigenvectorMatrix_1 \latentVector_{1}\right)\\ \latentVector_{3} &= \eigenvectwoMatrix^\top_3 \basisFunction\left(\eigenvectorMatrix_2 \latentVector_{2}\right)\\ \dataVector &= \mappingVector_4 ^\top \latentVector_{3} \end{align} $$

Now if we replace each of these neural networks with a Gaussian process. This is equivalent to taking the limit as the width of each layer goes to infinity, while appropriately scaling down the outputs.

$$ \begin{align} \latentVector_{1} &= \mappingFunctionVector_1\left(\inputVector\right)\\ \latentVector_{2} &= \mappingFunctionVector_2\left(\latentVector_{1}\right)\\ \latentVector_{3} &= \mappingFunctionVector_3\left(\latentVector_{2}\right)\\ \dataVector &= \mappingFunctionVector_4\left(\latentVector_{3}\right) \end{align} $$

The DeepFace architecture (Taigman et al. 2014) consists of layers that deal with translation and rotational invariances. These layers are followed by three locally-connected layers and two fully-connected layers. Color illustrates feature maps produced at each layer. The net includes more than 120 million parameters, where more than 95% come from the local and fully connected layers.

We can think of what these models are doing as being similar to early pin ball machines. In a neural network, we input a number (or numbers), whereas in pinball, we input a ball. The location of the ball on the left-right axis can be thought of as the number. As the ball falls through the machine, each layer of pins can be thought of as a different layer of neurons. Each layer acts to move the ball from left to right.

In a pinball machine, when the ball gets to the bottom it might fall into a hole defining a score, in a neural network, that is equivalent to the decision: a classification of the input object.

An image has more than one number associated with it, so it's like playing pinball in a hyper-space.

Learning involves moving all the pins to be in the right position, so that the ball falls in the right place. But moving all these pins in hyperspace can be difficult. In a hyper space you have to put a lot of data through the machine for to explore the positions of all the pins. Adversarial learning reflects the fact that a ball can be moved a small distance and lead to a very different result.

Probabilistic methods explore more of the space by considering a range of possible paths for the ball through the machine.

Mathematically, a deep Gaussian process can be seen as a composite multivariate function,
$$ \mathbf{g}(\inputVector)=\mappingFunctionVector_5(\mappingFunctionVector_4(\mappingFunctionVector_3(\mappingFunctionVector_2(\mappingFunctionVector_1(\inputVector))))). $$
Or if we view it from the probabilistic perspective we can see that a deep Gaussian process is specifying a factorization of the joint density, the standard deep model takes the form of a Markov chain.

from matplotlib import rc

rc("font", **{'family':'sans-serif','sans-serif':['Helvetica'],'size':30})
rc("text", usetex=True)

$$ p(\dataVector|\inputVector)= p(\dataVector|\mappingFunctionVector_5)p(\mappingFunctionVector_5|\mappingFunctionVector_4)p(\mappingFunctionVector_4|\mappingFunctionVector_3)p(\mappingFunctionVector_3|\mappingFunctionVector_2)p(\mappingFunctionVector_2|\mappingFunctionVector_1)p(\mappingFunctionVector_1|\inputVector) $$

Probabilistically the deep Gaussian process can be represented as a Markov chain.
from matplotlib import rc
rc("font", **{'family':'sans-serif','sans-serif':['Helvetica'], 'size':15})
rc("text", usetex=True)

Why Deep?

If the result of composing many functions together is simply another function, then why do we bother? The key point is that we can change the class of functions we are modeling by composing in this manner. A Gaussian process is specifying a prior over functions, and one with a number of elegant properties. For example, the derivative process (if it exists) of a Gaussian process is also Gaussian distributed. That makes it easy to assimilate, for example, derivative observations. But that also might raise some alarm bells. That implies that the marginal derivative distribution is also Gaussian distributed. If that's the case, then it means that functions which occasionally exhibit very large derivatives are hard to model with a Gaussian process. For example, a function with jumps in.

A one off discontinuity is easy to model with a Gaussian process, or even multiple discontinuities. They can be introduced in the mean function, or independence can be forced between two covariance functions that apply in different areas of the input space. But in these cases we will need to specify the number of discontinuities and where they occur. In otherwords we need to parameterise the discontinuities. If we do not know the number of discontinuities and don't wish to specify where they occur, i.e. if we want a non-parametric representation of discontinuities, then the standard Gaussian process doesn't help.

Stochastic Process Composition

The deep Gaussian process leads to non-Gaussian models, and non-Gaussian characteristics in the covariance function. In effect, what we are proposing is that we change the properties of the functions we are considering by *composing stochastic processes$. This is an approach to creating new stochastic processes from well known processes.

Additionally, we are not constrained to the formalism of the chain. For example, we can easily add single nodes emerging from some point in the depth of the chain. This allows us to combine the benefits of the graphical modelling formalism, but with a powerful framework for relating one set of variables to another, that of Gaussian processes

Difficulty for Probabilistic Approaches

The challenge for composition of probabilistic models is that you need to propagate a probability densities through non linear mappings. This allows you to create broader classes of probability density. Unfortunately it renders the resulting densities intractable.

The argument in the deep learning revolution is that deep architectures allow us to develop an abstraction of the feature set through model composition. Composing Gaussian processes is analytically intractable. To form deep Gaussian processes we use a variational approach to stack the models.

Stacked PCA

Composition of linear functions just leads to a new linear function.

Stacking a series of linear functions simply leads to a new linear function. The use of multiple linear function merely changes the covariance of the resulting Gaussian. If
$$ \latentMatrix \sim \gaussianSamp{\zerosVector}{\eye} $$
and the ith hidden layer is a multivariate linear transformation defined by $\weightMatrix_i$,
$$ \dataMatrix = \latentMatrix\weightMatrix_1 \weightMatrix_2 \dots \weightMatrix_\numLayers $$
then the rules of multivariate Gaussians tell us that
$$ \dataMatrix \sim \gaussianSamp{\zerosVector}{\weightMatrix_\numLayers \dots \weightMatrix_1 \weightMatrix^\top_1 \dots \weightMatrix^\top_\numLayers}. $$
So the model can be replaced by one where we set $\vMatrix = \weightMatrix_\numLayers \dots \weightMatrix_2 \weightMatrix_1$. So is such a model trivial? The answer is that it depends. There are two cases in which such a model remaisn interesting. Firstly, if we make intermediate observations stemming from the chain. So, for example, if we decide that,
$$ \latentMatrix_i = \weightMatrix_i \latentMatrix_{i-1} $$
and set $\latentMatrix_{0} = \inputMatrix \sim \gaussianSamp{\zerosVector}{\eye}$, then the matrices $\weightMatrix$ inter-relate a series of jointly Gaussian observations in an interesting way, stacking the full data matrix to give
$$ \latentMatrix = \begin{bmatrix} \latentMatrix_0 \\ \latentMatrix_1 \\ \vdots \\ \latentMatrix_\numLayers \end{bmatrix} $$
we can obtain
$$\latentMatrix \sim \gaussianSamp{\zerosVector}{\begin{bmatrix} \eye & \weightMatrix^\top_1 & \weightMatrix_1^\top\weightMatrix_2^\top & \dots & \vMatrix^\top \\ \weightMatrix_1 & \weightMatrix_1 \weightMatrix_1^\top & \weightMatrix_1 \weightMatrix_1^\top \weightMatrix_2^\top & \dots & \weightMatrix_1 \vMatrix^\top \\ \weightMatrix_2 \weightMatrix_1 & \weightMatrix_2 \weightMatrix_1 \weightMatrix_1^\top & \weightMatrix_2 \weightMatrix_1 \weightMatrix_1^\top \weightMatrix_2^\top & \dots & \weightMatrix_2 \weightMatrix_1 \vMatrix^\top \\ \vdots & \vdots & \vdots & \ddots & \vdots \\ \vMatrix & \vMatrix \weightMatrix_1^\top & \vMatrix \weightMatrix_1^\top \weightMatrix_2^\top& \dots & \vMatrix\vMatrix^\top \end{bmatrix}}$$
which is a highly structured Gaussian covariance with hierarchical dependencies between the variables $\latentMatrix_i$.

Stacked GP

Stacking Gaussian process models leads to non linear mappings at each stage. Here we are mapping from two dimensions to two dimensions in each layer.

Note that once the box has folded over on itself, it cannot be unfolded. So a feature that is generated near the top of the model cannot be removed furthr down the model.

This folding over effect happens in low dimensions. In higher dimensions it is less common.

Observation of this effect at a talk in Cambridge was one of the things that caused David Duvenaud (and collaborators) to consider the behavior of deeper Gaussian process models (Duvenaud et al. 2014).

Such folding over in the latent spaces necessarily forces the density to be non-Gaussian. Indeed, since folding-over is avoided as we increase the dimensionality of the latent spaces, such processes become more Gaussian. If we take the limit of the latent space dimensionality as it tends to infinity, the entire deep Gaussian process returns to a standard Gaussian process, with a covariance function given as a deep kernel (such as those described by Cho and Saul (2009)).

Further analysis of these deep networks has been conducted by Dunlop et al. (n.d.), who use analysis of the deep network's stationary density (treating it as a Markov chain across layers), to explore the nature of the implied process prior for a deep GP.

Both of these works, however, make constraining assumptions on the form of the Gaussian process prior at each layer (e.g. same covariance at each layer). In practice, the form of this covariance can be learnt and the densities described by the deep GP are more general than those mentioned in either of these papers.

David Duvenaud also created a YouTube video to help visualize what happens as you drop through the layers of a deep GP.

GPy: A Gaussian Process Framework in Python

GPy is a BSD licensed software code base for implementing Gaussian process models in python. This allows GPs to be combined with a wide variety of software libraries.

The software itself is avaialble on GitHub and the team welcomes contributions.

The aim for GPy is to be a probabilistic-style programming language, i.e. you specify the model rather than the algorithm. As well as a large range of covariance functions the software allows for non-Gaussian likelihoods, multivariate outputs, dimensionality reduction and approximations for larger data sets.

The GPy library can be installed via pip:

pip install GPy

This notebook depends on PyDeepGP. These libraries can be installed via pip:

pip install git+

Olympic Marathon Data

  • Gold medal times for Olympic Marathon since 1896.
  • Marathons before 1924 didn’t have a standardised distance.
  • Present results using pace per km.
  • In 1904 Marathon was badly organised leading to very slow times.
Image from Wikimedia Commons

The first thing we will do is load a standard data set for regression modelling. The data consists of the pace of Olympic Gold Medal Marathon winners for the Olympics from 1896 to present. First we load in the data and plot.

import numpy as np
import pods
data = pods.datasets.olympic_marathon_men()
x = data['X']
y = data['Y']

offset = y.mean()
scale = np.sqrt(y.var())

Things to notice about the data include the outlier in 1904, in this year, the olympics was in St Louis, USA. Organizational problems and challenges with dust kicked up by the cars following the race meant that participants got lost, and only very few participants completed.

More recent years see more consistently quick marathons.

Alan Turing

Alan Turing, in 1946 he was only 11 minutes slower than the winner of the 1948 games. Would he have won a hypothetical games held in 1946? Source: Alan Turing Internet Scrapbook.

If we had to summarise the objectives of machine learning in one word, a very good candidate for that word would be generalization. What is generalization? From a human perspective it might be summarised as the ability to take lessons learned in one domain and apply them to another domain. If we accept the definition given in the first session for machine learning,
$$ \text{data} + \text{model} \xrightarrow{\text{compute}} \text{prediction} $$
then we see that without a model we can't generalise: we only have data. Data is fine for answering very specific questions, like "Who won the Olympic Marathon in 2012?", because we have that answer stored, however, we are not given the answer to many other questions. For example, Alan Turing was a formidable marathon runner, in 1946 he ran a time 2 hours 46 minutes (just under four minutes per kilometer, faster than I and most of the other Endcliffe Park Run runners can do 5 km). What is the probability he would have won an Olympics if one had been held in 1946?

To answer this question we need to generalize, but before we formalize the concept of generalization let's introduce some formal representation of what it means to generalize in machine learning.

Our first objective will be to perform a Gaussian process fit to the data, we'll do this using the GPy software.

import GPy
m_full = GPy.models.GPRegression(x,yhat)
_ = m_full.optimize() # Optimize parameters of covariance function

The first command sets up the model, then m_full.optimize() optimizes the parameters of the covariance function and the noise level of the model. Once the fit is complete, we'll try creating some test points, and computing the output of the GP model in terms of the mean and standard deviation of the posterior functions between 1870 and 2030. We plot the mean function and the standard deviation at 200 locations. We can obtain the predictions using y_mean, y_var = m_full.predict(xt)

xt = np.linspace(1870,2030,200)[:,np.newaxis]
yt_mean, yt_var = m_full.predict(xt)

Now we plot the results using the helper function in teaching_plots.

Fit Quality

In the fit we see that the error bars (coming mainly from the noise variance) are quite large. This is likely due to the outlier point in 1904, ignoring that point we can see that a tighter fit is obtained. To see this making a version of the model, m_clean, where that point is removed.

x_clean=np.vstack((x[0:2, :], x[3:, :]))
y_clean=np.vstack((y[0:2, :], y[3:, :]))

m_clean = GPy.models.GPRegression(x_clean,y_clean)
_ = m_clean.optimize()

Deep GP Fit

Let's see if a deep Gaussian process can help here. We will construct a deep Gaussian process with one hidden layer (i.e. one Gaussian process feeding into another).

Build a Deep GP with an additional hidden layer (one dimensional) to fit the model.

import GPy
import deepgp
hidden = 1
m = deepgp.DeepGP([y.shape[1],hidden,x.shape[1]],Y=yhat, X=x, inits=['PCA','PCA'], 
                           GPy.kern.RBF(x.shape[1],ARD=True)], # the kernels for each layer
                  num_inducing=50, back_constraint=False)

Deep Gaussian process models also can require some thought in initialization. Here we choose to start by setting the noise variance to be one percent of the data variance.

Optimization requires moving variational parameters in the hidden layer representing the mean and variance of the expected values in that layer. Since all those values can be scaled up, and this only results in a downscaling in the output of the first GP, and a downscaling of the input length scale to the second GP. It makes sense to first of all fix the scales of the covariance function in each of the GPs.

Sometimes, deep Gaussian processes can find a local minima which involves increasing the noise level of one or more of the GPs. This often occurs because it allows a minimum in the KL divergence term in the lower bound on the likelihood. To avoid this minimum we habitually train with the likelihood variance (the noise on the output of the GP) fixed to some lower value for some iterations.

Let's create a helper function to initialize the models we use in the notebook.

import deepgp
from deepgp_tutorial import initialize
# Bind the new method to the Deep GP object.
# Call the initalization

Now optimize the model. The first stage of optimization is working on variational parameters and lengthscales only.


Now we remove the constraints on the scale of the covariance functions associated with each GP and optimize again.

for layer in m.layers:
    pass #layer.kern.variance.constrain_positive(warning=False)

Finally, we allow the noise variance to change and optimize for a large number of iterations.

for layer in m.layers:

For our optimization process we define a new function.

from deepgp_tutorial import staged_optimize
# Bind the new method to the Deep GP object.

Plot the prediction

The prediction of the deep GP can be extracted in a similar way to the normal GP. Although, in this case, it is an approximation to the true distribution, because the true distribution is not Gaussian.

import matplotlib.pyplot as plt
fig, ax = plt.subplots(figsize=plot.big_wide_figsize)
plot.model_output(m, scale=scale, offset=offset, ax=ax, xlabel='year', ylabel='pace min/km', 
          fontsize=20, portion=0.2)

mlai.write_figure(figure=fig, filename='../slides/diagrams/deepgp/olympic-marathon-deep-gp.svg', 
                transparent=True, frameon=True)

Olympic Marathon Data Deep GP

from deepgp_tutorial import posterior_sample
deepgp.DeepGP.posterior_sample = posterior_sample

Olympic Marathon Data Deep GP

Fitted GP for each layer

Now we explore the GPs the model has used to fit each layer. First of all, we look at the hidden layer.

from deepgp_tutorial import visualize
# Bind the new method to the Deep GP object.
from deepgp_tutorial import visualize_pinball
# Bind the new method to the Deep GP object.

Olympic Marathon Pinball Plot

The pinball plot shows the flow of any input ball through the deep Gaussian process. In a pinball plot a series of vertical parallel lines would indicate a purely linear function. For the olypmic marathon data we can see the first layer begins to shift from input towards the right. Note it also does so with some uncertainty (indicated by the shaded backgrounds). The second layer has less uncertainty, but bunches the inputs more strongly to the right. This input layer of uncertainty, followed by a layer that pushes inputs to the right is what gives the heteroschedastic noise.

Della Gatta Gene Data

  • Given given expression levels in the form of a time series from Della Gatta et al. (2008).
import numpy as np
import pods
data = pods.datasets.della_gatta_TRP63_gene_expression(data_set='della_gatta',gene_number=937)

x = data['X']
y = data['Y']

offset = y.mean()
scale = np.sqrt(y.var())
  • Want to detect if a gene is expressed or not, fit a GP to each gene Kalaitzis and Lawrence (2011).

Our first objective will be to perform a Gaussian process fit to the data, we'll do this using the GPy software.

import GPy
m_full = GPy.models.GPRegression(x,yhat)
_ = m_full.optimize() # Optimize parameters of covariance function

Initialize the length scale parameter (which here actually represents a time scale of the covariance function to a reasonable value. Default would be 1, but here we set it to 50 minutes, given points are arriving across zero to 250 minutes.

xt = np.linspace(-20,260,200)[:,np.newaxis]
yt_mean, yt_var = m_full.predict(xt)

Now we plot the results using the helper function in teaching_plots.

Now we try a model initialized with a longer length scale.

m_full2 = GPy.models.GPRegression(x,yhat)
_ = m_full2.optimize() # Optimize parameters of covariance function

Now we try a model initialized with a lower noise.

m_full3 = GPy.models.GPRegression(x,yhat)
_ = m_full3.optimize() # Optimize parameters of covariance function
layers = [y.shape[1], 1,x.shape[1]]
inits = ['PCA']*(len(layers)-1)
kernels = []
for i in layers[1:]:
    kernels += [GPy.kern.RBF(i)]
m = deepgp.DeepGP(layers,Y=yhat, X=x, 
                  kernels=kernels, # the kernels for each layer
                  num_inducing=20, back_constraint=False)

TP53 Gene Data Deep GP

TP53 Gene Data Deep GP

TP53 Gene Data Latent 1

TP53 Gene Data Latent 2

TP53 Gene Pinball Plot

Step Function

Next we consider a simple step function data set.

gap = -.1
x = np.vstack((np.linspace(-1, -gap/2.0, num_low)[:, np.newaxis],
              np.linspace(gap/2.0, 1, num_high)[:, np.newaxis]))
y = np.vstack((np.zeros((num_low, 1)), np.ones((num_high,1))))
scale = np.sqrt(y.var())
offset = y.mean()
yhat = (y-offset)/scale

Step Function Data

m_full = GPy.models.GPRegression(x,yhat)
_ = m_full.optimize() # Optimize parameters of covariance function

Step Function Data GP

layers = [y.shape[1], 1, 1, 1,x.shape[1]]
inits = ['PCA']*(len(layers)-1)
kernels = []
for i in layers[1:]:
    kernels += [GPy.kern.RBF(i)]
m = deepgp.DeepGP(layers,Y=yhat, X=x, 
                  kernels=kernels, # the kernels for each layer
                  num_inducing=20, back_constraint=False)

Step Function Data Deep GP

Step Function Data Deep GP

Step Function Data Latent 1

Step Function Data Latent 2

Step Function Data Latent 3

Step Function Data Latent 4

Step Function Pinball Plot

import pods
data = pods.datasets.mcycle()
x = data['X']
y = data['Y']
yhat = (y - offset)/scale

Motorcycle Helmet Data

m_full = GPy.models.GPRegression(x,yhat)
_ = m_full.optimize() # Optimize parameters of covariance function

Motorcycle Helmet Data GP

import deepgp
layers = [y.shape[1], 1, x.shape[1]]
inits = ['PCA']*(len(layers)-1)
kernels = []
for i in layers[1:]:
    kernels += [GPy.kern.RBF(i)]
m = deepgp.DeepGP(layers,Y=yhat, X=x, 
                  kernels=kernels, # the kernels for each layer
                  num_inducing=20, back_constraint=False)

m.staged_optimize(iters=(1000,1000,10000), messages=(True, True, True))

Motorcycle Helmet Data Deep GP

Motorcycle Helmet Data Deep GP

Motorcycle Helmet Data Latent 1

Motorcycle Helmet Data Latent 2

Motorcycle Helmet Pinball Plot

Robot Wireless Data

The robot wireless data is taken from an experiment run by Brian Ferris at University of Washington. It consists of the measurements of WiFi access point signal strengths as Brian walked in a loop.


x = np.linspace(0,1,215)[:, np.newaxis]
y = data['Y']
offset = y.mean()
scale = np.sqrt(y.var())
yhat = (y-offset)/scale

The ground truth is recorded in the data, the actual loop is given in the plot below.

Robot Wireless Ground Truth

We will ignore this ground truth in making our predictions, but see if the model can recover something similar in one of the latent layers.

Robot WiFi Data

Perform a Gaussian process fit on the data using GPy.

m_full = GPy.models.GPRegression(x,yhat)
_ = m_full.optimize() # Optimize parameters of covariance function

Robot WiFi Data GP

layers = [y.shape[1], 10, 5, 2, 2, x.shape[1]]
inits = ['PCA']*(len(layers)-1)
kernels = []
for i in layers[1:]:
    kernels += [GPy.kern.RBF(i, ARD=True)]
m = deepgp.DeepGP(layers,Y=y, X=x, inits=inits, 
                  num_inducing=50, back_constraint=False)

Robot WiFi Data Deep GP

Robot WiFi Data Deep GP

Robot WiFi Data Latent Space

Motion Capture

  • ‘High five’ data.
  • Model learns structure between two interacting subjects.

Shared LVM

Thanks to: Zhenwen Dai and Neil D. Lawrence

We now look at the deep Gaussian processes' capacity to perform unsupervised learning.

We will look at a sub-sample of the MNIST digit data set.

First load in the MNIST data set from scikit learn. This can take a little while because it's large to download.

from sklearn.datasets import fetch_mldata
mnist = fetch_mldata('MNIST original')

Sub-sample the dataset to make the training faster.

import numpy as np
digits = [0,1,2,3,4]
N_per_digit = 100
Y = []
labels = []
for d in digits:
    imgs = mnist['data'][mnist['target']==d]
Y = np.vstack(Y).astype(np.float64)
labels = np.hstack(labels)
Y /= 255.

Fit a Deep GP

We're going to fit a Deep Gaussian process model to the MNIST data with two hidden layers. Each of the two Gaussian processes (one from the first hidden layer to the second, one from the second hidden layer to the data) has an exponentiated quadratic covariance.

import deepgp
import GPy
num_latent = 2
num_hidden_2 = 5
m = deepgp.DeepGP([Y.shape[1],num_hidden_2,num_latent],
                  num_inducing=50, back_constraint=False, 


Just like deep neural networks, there are some tricks to intitializing these models. The tricks we use here include some early training of the model with model parameters constrained. This gives the variational inducing parameters some scope to tighten the bound for the case where the noise variance is small and the variances of the Gaussian processes are around 1.

m.obslayer.likelihood.variance[:] = Y.var()*0.01
for layer in m.layers:

We now we optimize for a hundred iterations with the constrained model.


Now we remove the fixed constraint on the kernel variance parameters, but keep the noise output constrained, and run for a further 100 iterations.

for layer in m.layers:

Finally we unconstrain the layer likelihoods and allow the full model to be trained for 1000 iterations.

for layer in m.layers:

Visualize the latent space of the top layer

Now the model is trained, let's plot the mean of the posterior distributions in the top latent layer of the model.

Visualize the latent space of the intermediate layer

We can also visualize dimensions of the intermediate layer. First the lengthscale of those dimensions is given by


Generate From Model

Now we can take a look at a sample from the model, by drawing a Gaussian random sample in the latent space and propagating it through the model.

rows = 10
cols = 20
t=np.linspace(-1, 1, rows*cols)[:, None]
kern = GPy.kern.RBF(1,lengthscale=0.05)
cov = kern.K(t, t)
x = np.random.multivariate_normal(np.zeros(rows*cols), cov, num_latent).T

  • Gaussian process based nonlinear latent structure discovery in multivariate spike train data Wu et al. (2017)
  • Doubly Stochastic Variational Inference for Deep Gaussian Processes Salimbeni and Deisenroth (2017)
  • Deep Multi-task Gaussian Processes for Survival Analysis with Competing Risks Alaa and van der Schaar (2017)
  • Counterfactual Gaussian Processes for Reliable Decision-making and What-if Reasoning Schulam and Saria (2017)

  • Deep Survival Analysis Ranganath et al. (2016)
  • Recurrent Gaussian Processes Mattos et al. (2015)
  • Gaussian Process Based Approaches for Survival Analysis A. D. Saul (2016)

Emukit Playground

Emukit playground is a software toolkit for exploring the use of statistical emulation as a tool. It was built by Adam Hirst, during his software engineering internship at Amazon and supervised by Cliff McCollum.

Emukit playground is a tutorial for understanding the simulation/emulation relationship.
Tutorial on Bayesian optimization of the number of taxis deployed from Emukit playground.

You can explore Bayesian optimization of a taxi simulation.

Uncertainty Quantification

Uncertainty quantification (UQ) is the science of quantitative characterization and reduction of uncertainties in both computational and real world applications. It tries to determine how likely certain outcomes are if some aspects of the system are not exactly known.

We will to illustrate different concepts of Uncertainty Quantification (UQ) and the role that Gaussian processes play in this field. Based on a simple simulator of a car moving between a valley and a mountain, we are going to illustrate the following concepts:

  • Systems emulation. Many real world decisions are based on simulations that can be computationally very demanding. We will show how simulators can be replaced by emulators: Gaussian process models fitted on a few simulations that can be used to replace the simulator. Emulators are cheap to compute, fast to run, and always provide ways to quantify the uncertainty of how precise they are compared the original simulator.

  • Emulators in optimization problems. We will show how emulators can be used to optimize black-box functions that are expensive to evaluate. This field is also called Bayesian Optimization and has gained an increasing relevance in machine learning as emulators can be used to optimize computer simulations (and machine learning algorithms) quite efficiently.

  • Multi-fidelity emulation methods. In many scenarios we have simulators of different quality about the same measure of interest. In these cases the goal is to merge all sources of information under the same model so the final emulator is cheaper and more accurate than an emulator fitted only using data from the most accurate and expensive simulator.

Mountain Car Simulator

To illustrate the above mentioned concepts we we use the mountain car simulator. This simulator is widely used in machine learning to test reinforcement learning algorithms. The goal is to define a control policy on a car whose objective is to climb a mountain. Graphically, the problem looks as follows:

The mountain car simulation from the Open AI gym.

The goal is to define a sequence of actions (push the car right or left with certain intensity) to make the car reach the flag after a number T of time steps.

At each time step t, the car is characterized by a vector $\inputVector_{t} = (p_t,v_t)$ of states which are respectively the the position and velocity of the car at time t. For a sequence of states (an episode), the dynamics of the car is given by

$$\inputVector_{t+1} = \mappingFunction(\inputVector_{t},\textbf{u}_{t})$$

where ut is the value of an action force, which in this example corresponds to push car to the left (negative value) or to the right (positive value). The actions across a full episode are represented in a policy $\textbf{u}_{t} = \pi(\inputVector_{t},\theta)$ that acts according to the current state of the car and some parameters θ. In the following examples we will assume that the policy is linear which allows us to write $\pi(\inputVector_{t},\theta)$ as

$$\pi(\inputVector,\theta)= \theta_0 + \theta_p p + \theta_vv.$$

For t = 1, …, T now given some initial state $\inputVector_{0}$ and some some values of each ut, we can simulate the full dynamics of the car for a full episode using Gym. The values of ut are fully determined by the parameters of the linear controller.

After each episode of length T is complete, a reward function RT(θ) is computed. In the mountain car example the reward is computed as 100 for reaching the target of the hill on the right hand side, minus the squared sum of actions (a real negative to push to the left and a real positive to push to the right) from start to goal. Note that our reward depend on θ as we make it dependent on the parameters of the linear controller.

Emulate the Mountain Car

import gym
env = gym.make('MountainCarContinuous-v0')

Our goal in this section is to find the parameters θ of the linear controller such that

θ* = argmaxθRT(θ).

In this section, we directly use Bayesian optimization to solve this problem. We will use GPyOpt so we first define the objective function:

import mountain_car as mc
import GPyOpt
obj_func = lambda x: mc.run_simulation(env, x)[0]
objective = GPyOpt.core.task.SingleObjective(obj_func)

For each set of parameter values of the linear controller we can run an episode of the simulator (that we fix to have a horizon of T = 500) to generate the reward. Using as input the parameters of the controller and as outputs the rewards we can build a Gaussian process emulator of the reward.

We start defining the input space, which is three-dimensional:

## --- We define the input space of the emulator

space= [{'name':'postion_parameter', 'type':'continuous', 'domain':(-1.2, +1)},
        {'name':'velocity_parameter', 'type':'continuous', 'domain':(-1/0.07, +1/0.07)},
        {'name':'constant', 'type':'continuous', 'domain':(-1, +1)}]

design_space = GPyOpt.Design_space(space=space)

Now we initizialize a Gaussian process emulator.

model = GPyOpt.models.GPModel(optimize_restarts=5, verbose=False, exact_feval=True, ARD=True)

In Bayesian optimization an acquisition function is used to balance exploration and exploitation to evaluate new locations close to the optimum of the objective. In this notebook we select the expected improvement (EI). For further details have a look to the review paper of Shahriari et al (2015).

aquisition_optimizer = GPyOpt.optimization.AcquisitionOptimizer(design_space)
acquisition = GPyOpt.acquisitions.AcquisitionEI(model, design_space, optimizer=aquisition_optimizer)
evaluator = GPyOpt.core.evaluators.Sequential(acquisition) # Collect points sequentially, no parallelization.

To initalize the model we start sampling some initial points (25) for the linear controler randomly.

from GPyOpt.experiment_design.random_design import RandomDesign
n_initial_points = 25
random_design = RandomDesign(design_space)
initial_design = random_design.get_samples(n_initial_points)

Before we start any optimization, lets have a look to the behavior of the car with the first of these initial points that we have selected randomly.

import numpy as np
random_controller = initial_design[0,:]
_, _, _, frames = mc.run_simulation(env, np.atleast_2d(random_controller), render=True)
anim=mc.animate_frames(frames, 'Random linear controller')
from IPython.core.display import HTML

As we can see the random linear controller does not manage to push the car to the top of the mountain. Now, let's optimize the regret using Bayesian optimization and the emulator for the reward. We try 50 new parameters chosen by the EI.

max_iter = 50
bo = GPyOpt.methods.ModularBayesianOptimization(model, design_space, objective, acquisition, evaluator, initial_design)
bo.run_optimization(max_iter = max_iter )

Now we visualize the result for the best controller that we have found with Bayesian optimization.

_, _, _, frames = mc.run_simulation(env, np.atleast_2d(bo.x_opt), render=True)
anim=mc.animate_frames(frames, 'Best controller after 50 iterations of Bayesian optimization')

he car can now make it to the top of the mountain! Emulating the reward function and using the EI helped as to find a linear controller that solves the problem.

Data Efficient Emulation

In the previous section we solved the mountain car problem by directly emulating the reward but no considerations about the dynamics $\inputVector_{t+1} = \mappingFunction(\inputVector_{t},\textbf{u}_{t})$ of the system were made. Note that we had to run 75 episodes of 500 steps each to solve the problem, which required to call the simulator 500 × 75 = 37500 times. In this section we will show how it is possible to reduce this number by building an emulator for f that can later be used to directly optimize the control.

The inputs of the model for the dynamics are the velocity, the position and the value of the control so create this space accordingly.

import gym
env = gym.make('MountainCarContinuous-v0')
import GPyOpt
space_dynamics = [{'name':'position', 'type':'continuous', 'domain':[-1.2, +0.6]},
                  {'name':'velocity', 'type':'continuous', 'domain':[-0.07, +0.07]},
                  {'name':'action', 'type':'continuous', 'domain':[-1, +1]}]
design_space_dynamics = GPyOpt.Design_space(space=space_dynamics)

The outputs are the velocity and the position. Indeed our model will capture the change in position and velocity on time. That is, we will model

Δvt + 1 = vt + 1 − vt

Δxt + 1 = pt + 1 − pt

with Gaussian processes with prior mean vt and pt respectively. As a covariance function, we use a Matern52. We need therefore two models to capture the full dynamics of the system.

position_model = GPyOpt.models.GPModel(optimize_restarts=5, verbose=False, exact_feval=True, ARD=True)
velocity_model = GPyOpt.models.GPModel(optimize_restarts=5, verbose=False, exact_feval=True, ARD=True)

Next, we sample some input parameters and use the simulator to compute the outputs. Note that in this case we are not running the full episodes, we are just using the simulator to compute $\inputVector_{t+1}$ given $\inputVector_{t}$ and ut.

import numpy as np
from GPyOpt.experiment_design.random_design import RandomDesign
import mountain_car as mc
### --- Random locations of the inputs
n_initial_points = 500
random_design_dynamics = RandomDesign(design_space_dynamics)
initial_design_dynamics = random_design_dynamics.get_samples(n_initial_points)
### --- Simulation of the (normalized) outputs
y = np.zeros((initial_design_dynamics.shape[0], 2))
for i in range(initial_design_dynamics.shape[0]):
    y[i, :] = mc.simulation(initial_design_dynamics[i, :])

# Normalize the data from the simulation
y_normalisation = np.std(y, axis=0)
y_normalised = y/y_normalisation

In general we might use much smarter strategies to design our emulation of the simulator. For example, we could use the variance of the predictive distributions of the models to collect points using uncertainty sampling, which will give us a better coverage of the space. For simplicity, we move ahead with the 500 randomly selected points.

Now that we have a data set, we can update the emulators for the location and the velocity.

position_model.updateModel(initial_design_dynamics, y[:, [0]], None, None)
velocity_model.updateModel(initial_design_dynamics, y[:, [1]], None, None)

We can now have a look to how the emulator and the simulator match. First, we show a contour plot of the car aceleration for each pair of can position and velocity. You can use the bar bellow to play with the values of the controler to compare the emulator and the simulator.

from IPython.html.widgets import interact

We can see how the emulator is doing a fairly good job approximating the simulator. On the edges, however, it struggles to captures the dynamics of the simulator.

Given some input parameters of the linear controlling, how do the dynamics of the emulator and simulator match? In the following figure we show the position and velocity of the car for the 500 time steps of an episode in which the parameters of the linear controller have been fixed beforehand. The value of the input control is also shown.

controller_gains = np.atleast_2d([0, .6, 1])  # change the valus of the linear controller to observe the trayectories.

We now make explicit use of the emulator, using it to replace the simulator and optimize the linear controller. Note that in this optimization, we don't need to query the simulator anymore as we can reproduce the full dynamics of an episode using the emulator. For illustrative purposes, in this example we fix the initial location of the car.

We define the objective reward function in terms of the simulator.

### --- Optimize control parameters with emulator
car_initial_location = np.asarray([-0.58912799, 0]) 

### --- Reward objective function using the emulator
obj_func_emulator = lambda x: mc.run_emulation([position_model, velocity_model], x, car_initial_location)[0]
objective_emulator = GPyOpt.core.task.SingleObjective(obj_func_emulator)

And as before, we use Bayesian optimization to find the best possible linear controller.

### --- Elements of the optimization that will use the multi-fidelity emulator
model = GPyOpt.models.GPModel(optimize_restarts=5, verbose=False, exact_feval=True, ARD=True)

The design space is the three continuous variables that make up the linear controller.

space= [{'name':'linear_1', 'type':'continuous', 'domain':(-1/1.2, +1)},
        {'name':'linear_2', 'type':'continuous', 'domain':(-1/0.07, +1/0.07)},
        {'name':'constant', 'type':'continuous', 'domain':(-1, +1)}]

design_space         = GPyOpt.Design_space(space=space)
aquisition_optimizer = GPyOpt.optimization.AcquisitionOptimizer(design_space)

random_design = RandomDesign(design_space)
initial_design = random_design.get_samples(25)

We set the acquisition function to be expected improvement using GPyOpt.

acquisition          = GPyOpt.acquisitions.AcquisitionEI(model, design_space, optimizer=aquisition_optimizer)
evaluator            = GPyOpt.core.evaluators.Sequential(acquisition)
bo_emulator = GPyOpt.methods.ModularBayesianOptimization(model, design_space, objective_emulator, acquisition, evaluator, initial_design)
_, _, _, frames = mc.run_simulation(env, np.atleast_2d(bo_emulator.x_opt), render=True)
anim=mc.animate_frames(frames, 'Best controller using the emulator of the dynamics')
from IPython.core.display import HTML

And the problem is again solved, but in this case we have replaced the simulator of the car dynamics by a Gaussian process emulator that we learned by calling the simulator only 500 times. Compared to the 37500 calls that we needed when applying Bayesian optimization directly on the simulator this is a great gain.

In some scenarios we have simulators of the same environment that have different fidelities, that is that reflect with different level of accuracy the dynamics of the real world. Running simulations of the different fidelities also have a different cost: hight fidelity simulations are more expensive the cheaper ones. If we have access to these simulators we can combine high and low fidelity simulations under the same model.

So let's assume that we have two simulators of the mountain car dynamics, one of high fidelity (the one we have used) and another one of low fidelity. The traditional approach to this form of multi-fidelity emulation is to assume that

$$\mappingFunction_i\left(\inputVector\right) = \rho\mappingFunction_{i-1}\left(\inputVector\right) + \delta_i\left(\inputVector \right)$$

where $\mappingFunction_{i-1}\left(\inputVector\right)$ is a low fidelity simulation of the problem of interest and $\mappingFunction_i\left(\inputVector\right)$ is a higher fidelity simulation. The function $\delta_i\left(\inputVector \right)$ represents the difference between the lower and higher fidelity simulation, which is considered additive. The additive form of this covariance means that if $\mappingFunction_{0}\left(\inputVector\right)$ and $\left\{\delta_i\left(\inputVector \right)\right\}_{i=1}^m$ are all Gaussian processes, then the process over all fidelities of simuation will be a joint Gaussian process.

But with Deep Gaussian processes we can consider the form

$$\mappingFunction_i\left(\inputVector\right) = \mappingFunctionTwo_{i}\left(\mappingFunction_{i-1}\left(\inputVector\right)\right) + \delta_i\left(\inputVector \right),$$

where the low fidelity representation is non linearly transformed by $\mappingFunctionTwo(\cdot)$ before use in the process. This is the approach taken in Perdikaris et al. (2017). But once we accept that these models can be composed, a highly flexible framework can emerge. A key point is that the data enters the model at different levels, and represents different aspects. For example these correspond to the two fidelities of the mountain car simulator.

We start by sampling both of them at 250 random input locations.

import gym
env = gym.make('MountainCarContinuous-v0')
import GPyOpt
### --- Collect points from low and high fidelity simulator --- ###

space = GPyOpt.Design_space([
        {'name':'position', 'type':'continuous', 'domain':(-1.2, +1)},
        {'name':'velocity', 'type':'continuous', 'domain':(-0.07, +0.07)},
        {'name':'action', 'type':'continuous', 'domain':(-1, +1)}])

n_points = 250
random_design = GPyOpt.experiment_design.RandomDesign(space)
x_random = random_design.get_samples(n_points)

Next, we evaluate the high and low fidelity simualtors at those locations.

import numpy as np
import mountain_car as mc
d_position_hf = np.zeros((n_points, 1))
d_velocity_hf = np.zeros((n_points, 1))
d_position_lf = np.zeros((n_points, 1))
d_velocity_lf = np.zeros((n_points, 1))

# --- Collect high fidelity points
for i in range(0, n_points):
    d_position_hf[i], d_velocity_hf[i] = mc.simulation(x_random[i, :])

# --- Collect low fidelity points  
for i in range(0, n_points):
    d_position_lf[i], d_velocity_lf[i] = mc.low_cost_simulation(x_random[i, :])

It is time to build the multi-fidelity model for both the position and the velocity.

As we did in the previous section we use the emulator to optimize the simulator. In this case we use the high fidelity output of the emulator.

### --- Optimize controller parameters 
obj_func = lambda x: mc.run_simulation(env, x)[0]
obj_func_emulator = lambda x: mc.run_emulation([position_model, velocity_model], x, car_initial_location)[0]
objective_multifidelity = GPyOpt.core.task.SingleObjective(obj_func)

And we optimize using Bayesian optimzation.

from GPyOpt.experiment_design.random_design import RandomDesign
model = GPyOpt.models.GPModel(optimize_restarts=5, verbose=False, exact_feval=True, ARD=True)
space= [{'name':'linear_1', 'type':'continuous', 'domain':(-1/1.2, +1)},
        {'name':'linear_2', 'type':'continuous', 'domain':(-1/0.07, +1/0.07)},
        {'name':'constant', 'type':'continuous', 'domain':(-1, +1)}]

design_space = GPyOpt.Design_space(space=space)
aquisition_optimizer = GPyOpt.optimization.AcquisitionOptimizer(design_space)

n_initial_points = 25
random_design = RandomDesign(design_space)
initial_design = random_design.get_samples(n_initial_points)
acquisition = GPyOpt.acquisitions.AcquisitionEI(model, design_space, optimizer=aquisition_optimizer)
evaluator = GPyOpt.core.evaluators.Sequential(acquisition)
bo_multifidelity = GPyOpt.methods.ModularBayesianOptimization(model, design_space, objective_multifidelity, acquisition, evaluator, initial_design)
_, _, _, frames = mc.run_simulation(env, np.atleast_2d(bo_multifidelity.x_opt), render=True)
anim=mc.animate_frames(frames, 'Best controller with multi-fidelity emulator')
from IPython.core.display import HTML

And problem solved! We see how the problem is also solved with 250 observations of the high fidelity simulator and 250 of the low fidelity simulator.


  • Multi-fidelity emulation: build surrogate models when data is obtained from multiple information sources that have different fidelity and/or cost;
  • Bayesian optimisation: optimise physical experiments and tune parameters of machine learning algorithms;
  • Experimental design/Active learning: design the most informative experiments and perform active learning with machine learning models;
  • Sensitivity analysis: analyse the influence of inputs on the outputs of a given system;
  • Bayesian quadrature: efficiently compute the integrals of functions that are expensive to evaluate.

MXFusion: Modular Probabilistic Programming on MXNet
</tr> </table>


Why another framework?

Key Requirements

Specialized inference methods + models, without requiring users to reimplement nor understand them every time. Leverage expert knowledge. Efficient inference, flexible framework. Existing frameworks either did one or the other: flexible, or efficient.

What does it look like?



m = Model() = Variable()
m.s = Variable(transformation=PositiveTransformation())
m.Y = Normal.define_variable(, variance=m.s)
  • Variable
  • Distribution
  • Function

  • log_pdf
  • draw_samples

  • Variational Inference
  • MCMC Sampling (soon) Built on MXNet Gluon (imperative code, not static graph)

infr = GradBasedInference(inference_algorithm=MAP(model=m, observed=[m.Y]))
  • Model + Inference together form building blocks.
    • Just doing modular modeling with universal inference doesn't really scale, need specialized inference methods for specialized modelling objects like non-parametrics.

Long term Aim

  • Simulate/Emulate the components of the system.
    • Validate with real world using multifidelity.
    • Interpret system using e.g. sensitivity analysis.
  • Perform end to end learning to optimize.
    • Maintain interpretability.

Stefanos Eleftheriadis, John Bronskill, Hugh Salimbeni, Rich Turner, Zhenwen Dai, Javier Gonzalez, Andreas Damianou, Mark Pullin, Eric Meissner.


Alaa, Ahmed M., and Mihaela van der Schaar. 2017. “Deep Multi-Task Gaussian Processes for Survival Analysis with Competing Risks.” In Advances in Neural Information Processing Systems 30, edited by I. Guyon, U. V. Luxburg, S. Bengio, H. Wallach, R. Fergus, S. Vishwanathan, and R. Garnett, 2326–34. Curran Associates, Inc.

Álvarez, Mauricio A., David Luengo, Michalis K. Titsias, and Neil D. Lawrence. 2010. “Efficient Multioutput Gaussian Processes Through Variational Inducing Kernels.” In, 9:25–32.

Bui, Thang D., Josiah Yan, and Richard E. Turner. 2017. “A Unifying Framework for Gaussian Process Pseudo-Point Approximations Using Power Expectation Propagation.” Journal of Machine Learning Research 18 (104): 1–72.

Cho, Youngmin, and Lawrence K. Saul. 2009. “Kernel Methods for Deep Learning.” In Advances in Neural Information Processing Systems 22, edited by Y. Bengio, D. Schuurmans, J. D. Lafferty, C. K. I. Williams, and A. Culotta, 342–50. Curran Associates, Inc.

Csató, Lehel. 2002. “Gaussian Processes — Iterative Sparse Approximations.” PhD thesis, Aston University.

Csató, Lehel, and Manfred Opper. 2002. “Sparse on-Line Gaussian Processes.” Neural Computation 14 (3): 641–68.

Dai, Zhenwen, Andreas Damianou, James Hensman, and Neil D. Lawrence. 2014. “Gaussian Process Models with Parallelization and GPU Acceleration.”

Damianou, Andreas. 2015. “Deep Gaussian Processes and Variational Propagation of Uncertainty.” PhD thesis, University of Sheffield.

Della Gatta, Giusy, Mukesh Bansal, Alberto Ambesi-Impiombato, Dario Antonini, Caterina Missero, and Diego di Bernardo. 2008. “Direct Targets of the Trp63 Transcription Factor Revealed by a Combination of Gene Expression Profiling and Reverse Engineering.” Genome Research 18 (6). Telethon Institute of Genetics; Medicine, 80131 Naples, Italy.: 939–48. doi:10.1101/gr.073601.107.

Dunlop, Matthew M., Mark A. Girolami, Andrew M. Stuart, and Aretha L. Teckentrup. n.d. “How Deep Are Deep Gaussian Processes?” Journal of Machine Learning Research 19 (54): 1–46.

Duvenaud, David, Oren Rippel, Ryan Adams, and Zoubin Ghahramani. 2014. “Avoiding Pathologies in Very Deep Networks.” In.

Gal, Yarin, Mark van der Wilk, and Carl E. Rasmussen. n.d. “Distributed Variational Inference in Sparse Gaussian Process Regression and Latent Variable Models.” In.

Hensman, James, Nicoló Fusi, and Neil D. Lawrence. n.d. “Gaussian Processes for Big Data.” In.

Hoffman, Matthew, David M. Blei, Chong Wang, and John Paisley. 2012. “Stochastic Variational Inference.” arXiv Preprint arXiv:1206.7051.

Kalaitzis, Alfredo A., and Neil D. Lawrence. 2011. “A Simple Approach to Ranking Differentially Expressed Gene Expression Time Courses Through Gaussian Process Regression.” BMC Bioinformatics 12 (180). doi:10.1186/1471-2105-12-180.

Lawrence, Neil D. n.d. “Learning for Larger Datasets with the Gaussian Process Latent Variable Model.” In, 243–50.

Lázaro-Gredilla, Miguel, Joaquin Quiñonero-Candela, and Carl Edward Rasmussen. 2010. “Sparse Spectrum Gaussian Processes.” Journal of Machine Learning Research 11: 1865–81.

MacKay, David J. C. n.d. “Introduction to Gaussian Processes.” In, 133–66.

Mattos, César Lincoln C., Zhenwen Dai, Andreas C. Damianou, Jeremy Forth, Guilherme A. Barreto, and Neil D. Lawrence. 2015. “Recurrent Gaussian Processes.” CoRR abs/1511.06644.

Perdikaris, P., M. Raissi, A. Damianou, N. D. Lawrence, and G. E. Karniadakis. 2017. “Nonlinear Information Fusion Algorithms for Data-Efficient Multi-Fidelity Modelling.” Proceedings of the Royal Society of London A: Mathematical, Physical and Engineering Sciences 473 (2198). The Royal Society. doi:10.1098/rspa.2016.0751.

Quiñonero Candela, Joaquin, and Carl Edward Rasmussen. 2005. “A Unifying View of Sparse Approximate Gaussian Process Regression.” Journal of Machine Learning Research 6: 1939–59.

Ranganath, Rajesh, Adler Perotte, Noémie Elhadad, and David Blei. 2016. “Deep Survival Analysis.” In Proceedings of the 1st Machine Learning for Healthcare Conference, edited by Finale Doshi-Velez, Jim Fackler, David Kale, Byron Wallace, and Jenna Wiens, 56:101–14. Proceedings of Machine Learning Research. Children’s Hospital LA, Los Angeles, CA, USA: PMLR.

Salimbeni, Hugh, and Marc Deisenroth. 2017. “Doubly Stochastic Variational Inference for Deep Gaussian Processes.” In Advances in Neural Information Processing Systems 30, edited by I. Guyon, U. V. Luxburg, S. Bengio, H. Wallach, R. Fergus, S. Vishwanathan, and R. Garnett, 4591–4602. Curran Associates, Inc.

Saul, Alan Daniel. 2016. “Gaussian Process Based Approaches for Survival Analysis.” PhD thesis, University of Sheffield.

Schulam, Peter, and Suchi Saria. 2017. “Counterfactual Gaussian Processes for Reliable Decision-Making and What-If Reasoning.” In Advances in Neural Information Processing Systems 30, edited by I. Guyon, U. V. Luxburg, S. Bengio, H. Wallach, R. Fergus, S. Vishwanathan, and R. Garnett, 1696–1706. Curran Associates, Inc.

Seeger, Matthias W., Asmus Hetzel, Zhenwen Dai, and Neil D. Lawrence. 2017. “Auto-Differentiating Linear Algebra.” CoRR abs/1710.08717.

Seeger, Matthias, Christopher K. I. Williams, and Neil D. Lawrence. n.d. “Fast Forward Selection to Speed up Sparse Gaussian Process Regression.” In.

Smola, Alexander J., and Peter L. Bartlett. n.d. “Sparse Greedy Gaussian Process Regression.” In, 619–25.

Snelson, Edward, and Zoubin Ghahramani. n.d. “Sparse Gaussian Processes Using Pseudo-Inputs.” In.

Taigman, Yaniv, Ming Yang, Marc’Aurelio Ranzato, and Lior Wolf. 2014. “DeepFace: Closing the Gap to Human-Level Performance in Face Verification.” In Proceedings of the IEEE Computer Society Conference on Computer Vision and Pattern Recognition. doi:10.1109/CVPR.2014.220.

Titsias, Michalis K. n.d. “Variational Learning of Inducing Variables in Sparse Gaussian Processes.” In, 567–74.

Williams, Christopher K. I., and Matthias Seeger. n.d. “Using the Nyström Method to Speed up Kernel Machines.” In, 682–88.

Wu, Anqi, Nicholas G Roy, Stephen Keeley, and Jonathan W Pillow. 2017. “Gaussian Process Based Nonlinear Latent Structure Discovery in Multivariate Spike Train Data.” In Advances in Neural Information Processing Systems 30, edited by I. Guyon, U. V. Luxburg, S. Bengio, H. Wallach, R. Fergus, S. Vishwanathan, and R. Garnett, 3499–3508. Curran Associates, Inc.

  • Work by Eric Meissner and Zhenwen Dai.
  • Probabilistic programming.
  • Available on Github </td>