Abstract

In this talk we introduce deep Gaussian processes, an approach to stochastic process modelling that relies on the composition of individual stochastic proceses.

Deep Gaussian Processes

# Download some utilty files

import urllib.request

urllib.request.urlretrieve('https://raw.githubusercontent.com/lawrennd/talks/gh-pages/teaching_plots.py','teaching_plots.py')

urllib.request.urlretrieve('https://raw.githubusercontent.com/lawrennd/talks/gh-pages/mlai.py','mlai.py')

urllib.request.urlretrieve('https://raw.githubusercontent.com/lawrennd/talks/gh-pages/gp_tutorial.py','gp_tutorial.py')

urllib.request.urlretrieve('https://raw.githubusercontent.com/lawrennd/talks/gh-pages/deepgp_tutorial.py','deepgp_tutorial.py')

import os
for path in ['gp', 'datasets', 'deepgp']:
    if not os.path.exists(path):
        os.mkdir(path)

%pip install --upgrade git+https://github.com/sods/ods

%pip install gpy

GPy: A Gaussian Process Framework in Python

[edit]

Gaussian processes are a flexible tool for non-parametric analysis with uncertainty. The GPy software was started in Sheffield to provide a easy to use interface to GPs. One which allowed the user to focus on the modelling rather than the mathematics.

Figure: GPy is a BSD licensed software code base for implementing Gaussian process models in Python. It is designed for teaching and modelling. We welcome contributions which can be made through the Github repository https://github.com/SheffieldML/GPy

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 available 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 documentation for GPy can be found here.

This notebook depends on PyDeepGP. This library can be installed via pip.

%pip install --upgrade git+https://github.com/SheffieldML/PyDeepGP.git

Universe isn’t as Gaussian as it Was

[edit]

The Planck space craft was a European Space Agency space telescope that mapped the cosmic microwave background (CMB) from 2009 to 2013. The Cosmic Microwave Background is the first observable echo we have of the big bang. It dates to approximately 400,000 years after the big bang, at the time the universe was approximately 10⁸ times smaller and the temperature of the Univers was high, around 3 × 10⁸ degrees Kelvin. The Universe was in the form of a hydrogen plasma. The echo we observe is the moment when the Universe was cool enough for Protons and electrons to combine to form hydrogen atoms. At this moment, the Universe became transparent for the first time, and photons could travel through space.

Figure: Artist’s impression of the Planck spacecraft which measured the Cosmic Microwave Background between 2009 and 2013.

The objective of the Planck space craft was to measure the anisotropy and statistics of the Cosmic Microwave Background. This was important, because if the standard model of the Universe is correct the variations around the very high temperature of the Universe of the CMB should be distributed according to a Gaussian process.¹ Currently our best estimates show this to be the case (Jaffe et al. 1998; Pontzen and Peiris 2010; Elsner, Leistedt, and Peiris 2015, 2016).

To the high degree of precision that we could measure with the Planck space telescope, the CMB appears to be a Gaussian process. The parameters of its covariance function are given by the fundamental parameters of the universe, for example the amount of dark matter and matter in the universe

Figure: The cosmic microwave background is, to a very high degree of precision, a Gaussian process. The parameters of its covariance function are given by fundamental parameters of the universe, such as the amount of dark matter and mass.

Simulating a CMB Map

The simulation was created by Boris Leistedt, see the original Jupter notebook here.

Here we use that code to simulate our own universe and sample from what it looks like.

First we install some specialist software as well as matplotlib, scipy, numpy we require

• camb: http://camb.readthedocs.io/en/latest/
• healpy: https://healpy.readthedocs.io/en/latest/

%pip install camb

%pip install healpy

%matplotlib inline
%config IPython.matplotlib.backend = 'retina'
%config InlineBackend.figure_format = 'retina'

import matplotlib
import matplotlib.pyplot as plt
from matplotlib import rc
from cycler import cycler

import numpy as np

rc("font", family="serif", size=14)
rc("text", usetex=False)
matplotlib.rcParams['lines.linewidth'] = 2
matplotlib.rcParams['patch.linewidth'] = 2
matplotlib.rcParams['axes.prop_cycle'] =\
    cycler("color", ['k', 'c', 'm', 'y'])
matplotlib.rcParams['axes.labelsize'] = 16

import healpy as hp

import camb
from camb import model, initialpower

Now we use the theoretical power spectrum to design the covariance function.

nside = 512  # Healpix parameter, giving 12*nside**2 equal-area pixels on the sphere.
lmax = 3*nside # band-limit. Should be 2*nside < lmax < 4*nside to get information content.

Now we design our Universe. It is parameterised according to the ΛCDM model. The variables are as follows. H0 is the Hubble parameter (in Km/s/Mpc). The ombh2 is Physical Baryon density parameter. The omch2 is the physical dark matter density parameter. mnu is the sum of the neutrino masses (in electron Volts). omk is the Ω_k is the curvature parameter, which is here set to 0, tiving the minimal six parameter Lambda-CDM model. tau is the reionization optical depth.

Then we set ns, the “scalar spectral index”. This was estimated by Planck to be 0.96. Then there’s r, the ratio of the tensor power spectrum to scalar power spectrum. This has been estimated by Planck to be under 0.11. Here we set it to zero. These parameters are associated with inflation.

# Mostly following http://camb.readthedocs.io/en/latest/CAMBdemo.html with parameters from https://en.wikipedia.org/wiki/Lambda-CDM_model

pars = camb.CAMBparams()
pars.set_cosmology(H0=67.74, ombh2=0.0223, omch2=0.1188, mnu=0.06, omk=0, tau=0.066)
pars.InitPower.set_params(ns=0.96, r=0)

Having set the parameters, we now use the python software “Code for Anisotropies in the Microwave Background” to get the results.

pars.set_for_lmax(lmax, lens_potential_accuracy=0);
results = camb.get_results(pars)
powers = results.get_cmb_power_spectra(pars)
totCL = powers['total']
unlensedCL = powers['unlensed_scalar']

ells = np.arange(totCL.shape[0])
Dells = totCL[:, 0]
Cells = Dells * 2*np.pi / ells / (ells + 1)  # change of convention to get C_ell
Cells[0:2] = 0

cmbmap = hp.synfast(Cells, nside, 
                 lmax=lmax, mmax=None, alm=False, pol=False, 
                 pixwin=False, fwhm=0.0, sigma=None, new=False, verbose=True)

Figure: A simulation of the Cosmic Microwave Background obtained through sampling from the relevant Gaussian process covariance (in polar co-ordinates).

The world we see today, of course, is not a Gaussian process. There are many dicontinuities, for example, in the density of matter, and therefore in the temperature of the Universe.

 = f()

Figure: What we observe today is some non-linear function of the cosmic microwave background.

We can think of todays observed Universe, though, as a being a consequence of those temperature fluctuations in the CMB. Those fluctuations are only order 10⁻6 of the scale of the overal temperature of the Universe. But minor fluctations in that density is what triggered the pattern of formation of the Galaxies and how stars formed and created the elements that are the building blocks of our Earth (Vogelsberger et al. 2020).

Low Rank Gaussian Processes

[edit]

Figure: In recent years, approximations for Gaussian process models haven’t been the most fashionable approach to machine learning. 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 Nyström 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

[edit]

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), 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

[edit]

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 $p(\dataVector |\inducingVector)$

[edit]

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 (Gal, Wilk, and Rasmussen, n.d.; Dai et al. 2014; Seeger et al. 2017).

A Simple Regression Problem

[edit]

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

np.random.seed(101)

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

Figure: 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(
        (np.linspace(2.5,4.,3),
        np.linspace(7,8.5,3)))[:,None]
m = GPy.models.SparseGPRegression(X,y,kernel=kern,Z=Z)
m.noise_var = noise_var
m.inducing_inputs.constrain_fixed()
display(m)

Figure: Sparse Gaussian process fitted with six inducing variables, no optimization of parameters or inducing variables.

_ = m.optimize(messages=True)
display(m)

Figure: Gaussian process fitted with inducing variables fixed and parameters optimized

m.randomize()
m.inducing_inputs.unconstrain()
_ = m.optimize(messages=True)

Figure: 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.num_inducing=8
m.randomize()
M = 8
m.set_Z(np.random.rand(M,1)*12)

_ = m.optimize(messages=True)

Figure: 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())

Modern Review

• 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.)

Figure: 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

[edit]

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 k_i − 1 nodes in one layer, and k_i nodes in the following, then that matrix contains k_ik_i − 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.
$$ \mappingMatrix = \eigenvectorMatrix\eigenvalueMatrix\eigenvectwoMatrix^\top $$
or
$$ \mappingMatrix = \eigenvectorMatrix\eigenvectwoMatrix^\top $$
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.

Figure: Pictorial representation of the low rank form of the matrix $\mappingMatrix$.

Bottleneck Layers in Deep Neural Networks

[edit]

Figure: Inserting the bottleneck layers introduces a new set of variables.

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
$$ \begin{align} \latentVector_{1} &= \eigenvectwoMatrix^\top_1 \inputVector\\ \hiddenVector_{1} &= \basisFunction\left(\eigenvectorMatrix_1 \latentVector_{1}\right)\\ \latentVector_{2} &= \eigenvectwoMatrix^\top_2 \hiddenVector_{1}\\ \hiddenVector_{2} &= \basisFunction\left(\eigenvectorMatrix_2 \latentVector_{2}\right)\\ \latentVector_{3} &= \eigenvectwoMatrix^\top_3 \hiddenVector_{2}\\ \hiddenVector_{3} &= \basisFunction\left(\eigenvectorMatrix_3 \latentVector_{3}\right)\\ \dataVector &= \mappingVector_4^\top\hiddenVector_{3}. \end{align} $$

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

Cascade of Gaussian Processes

[edit]

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.

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

Deep Learning

[edit]

DeepFace

[edit]

Figure: The DeepFace architecture (Taigman et al. 2014), visualized through colors to represent the functional mappings at each layer. There are 120 million parameters in the model.

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 neural network includes more than 120 million parameters, where more than 95% come from the local and fully connected layers.

Deep Learning as Pinball

[edit]

Figure: Deep learning models are composition of simple functions. We can think of a pinball machine as an analogy. Each layer of pins corresponds to one of the layers of functions in the model. Input data is represented by the location of the ball from left to right when it is dropped in from the top. Output class comes from the position of the ball as it leaves the pins at the bottom.

Sometimes deep learning models are described as being like the brain, or too complex to understand, but one analogy I find useful to help the gist of these models is to think of them as being similar to early pin ball machines.

In a deep neural network, we input a number (or numbers), whereas in pinball, we input a ball.

Think of the location of the ball on the left-right axis as a single number. Our simple pinball machine can only take one number at a time. 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 is like playing pinball in a hyper-space.

Figure: At initialization, the pins, which represent the parameters of the function, aren’t in the right place to bring the balls to the correct decisions.

Figure: After learning the pins are now in the right place to bring the balls to the correct decisions.

Learning involves moving all the pins to be in the correct position, so that the ball ends up in the right place when it’s fallen through the machine. 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. Even when you feed many millions of data points through the machine, there are likely to be regions in the hyper-space where no ball has passed. When future test data passes through the machine in a new route unusual things can happen.

Adversarial examples exploit this high dimensional space. If you have access to the pinball machine, you can use gradient methods to find a position for the ball in the hyper space where the image looks like one thing, but will be classified as another.

Probabilistic methods explore more of the space by considering a range of possible paths for the ball through the machine. This helps to make them more data efficient and gives some robustness to adversarial examples.

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.

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

Figure: Probabilistically the deep Gaussian process can be represented as a Markov chain.

Figure: More usually deep probabilistic models are written vertically rather than horizontally as in the Markov chain.

Why Deep?

[edit]

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

Figure: More generally we aren’t constrained by the Markov chain. We can design structures that respect our belief about the underlying conditional dependencies. Here we are adding a side note from the chain.

Difficulty for Probabilistic Approaches

[edit]

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.

Figure: A two dimensional grid mapped into three dimensions to form a two dimensional manifold.

Figure: A one dimensional line mapped into two dimensions by two separate independent functions. Each point can be mapped exactly through the mappings.

Figure: A Gaussian density over the input of a non linear function leads to a very non Gaussian output. Here the output is multimodal.

Standard Variational Approach Fails

[edit]

• Standard variational bound has the form:
  $$ \likelihoodBound = \expDist{\log p(\dataVector|\latentMatrix)}{q(\latentMatrix)} + \KL{q(\latentMatrix)}{p(\latentMatrix)} $$
  

The standard variational approach would require the expectation of $\log p(\dataVector|\latentMatrix)$ under $q(\latentMatrix)$.
$$ \begin{align} \log p(\dataVector|\latentMatrix) = & -\frac{1}{2}\dataVector^\top\left(\kernelMatrix_{\mappingFunctionVector, \mappingFunctionVector}+\dataStd^2\eye\right)^{-1}\dataVector \\ & -\frac{1}{2}\log \det{\kernelMatrix_{\mappingFunctionVector, \mappingFunctionVector}+\dataStd^2 \eye} -\frac{\numData}{2}\log 2\pi \end{align} $$
But this is extremely difficult to compute because $\kernelMatrix_{\mappingFunctionVector, \mappingFunctionVector}$ is dependent on $\latentMatrix$ and it appears in the inverse.

Variational Bayesian GP-LVM

The alternative approach is to consider the collapsed variational bound (used for low rank (sparse is a misnomer) Gaussian process approximations.
$$ p(\dataVector)\geq \prod_{i=1}^\numData c_i \int \gaussianDist{\dataVector}{\expSamp{\mappingFunctionVector}}{\dataStd^2\eye}p(\inducingVector) \text{d}\inducingVector $$

$$ p(\dataVector|\latentMatrix )\geq \prod_{i=1}^\numData c_i \int \gaussianDist{\dataVector}{\expDist{\mappingFunctionVector}{p(\mappingFunctionVector|\inducingVector, \latentMatrix)}}{\dataStd^2\eye}p(\inducingVector) \text{d}\inducingVector $$

$$ \int p(\dataVector|\latentMatrix)p(\latentMatrix) \text{d}\latentMatrix \geq \int \prod_{i=1}^\numData c_i \gaussianDist{\dataVector}{\expDist{\mappingFunctionVector}{p(\mappingFunctionVector|\inducingVector, \latentMatrix)}}{\dataStd^2\eye} p(\latentMatrix)\text{d}\latentMatrix p(\inducingVector) \text{d}\inducingVector $$

To integrate across $\latentMatrix$ we apply the lower bound to the inner integral.
$$ \begin{align} \int \prod_{i=1}^\numData c_i \gaussianDist{\dataVector}{\expDist{\mappingFunctionVector}{p(\mappingFunctionVector|\inducingVector, \latentMatrix)}}{\dataStd^2\eye} p(\latentMatrix)\text{d}\latentMatrix \geq & \expDist{\sum_{i=1}^\numData\log c_i}{q(\latentMatrix)}\\ & +\expDist{\log\gaussianDist{\dataVector}{\expDist{\mappingFunctionVector}{p(\mappingFunctionVector|\inducingVector, \latentMatrix)}}{\dataStd^2\eye}}{q(\latentMatrix)}\\& + \KL{q(\latentMatrix)}{p(\latentMatrix)} \end{align} $$
* Which is analytically tractable for Gaussian $q(\latentMatrix)$ and some covariance functions.

• Need expectations under $q(\latentMatrix)$ of:
  $$ \log c_i = \frac{1}{2\dataStd^2} \left[\kernelScalar_{i, i} - \kernelVector_{i, \inducingVector}^\top \kernelMatrix_{\inducingVector, \inducingVector}^{-1} \kernelVector_{i, \inducingVector}\right] $$
  and
  $$ \log \gaussianDist{\dataVector}{\expDist{\mappingFunctionVector}{p(\mappingFunctionVector|\inducingVector,\dataMatrix)}}{\dataStd^2\eye} = -\frac{1}{2}\log 2\pi\dataStd^2 - \frac{1}{2\dataStd^2}\left(\dataScalar_i - \kernelMatrix_{\mappingFunctionVector, \inducingVector}\kernelMatrix_{\inducingVector,\inducingVector}^{-1}\inducingVector\right)^2 $$
  

• This requires the expectations
  $$ \expDist{\kernelMatrix_{\mappingFunctionVector,\inducingVector}}{q(\latentMatrix)} $$
  and
  $$ \expDist{\kernelMatrix_{\mappingFunctionVector,\inducingVector}\kernelMatrix_{\inducingVector,\inducingVector}^{-1}\kernelMatrix_{\inducingVector,\mappingFunctionVector}}{q(\latentMatrix)} $$
  which can be computed analytically for some covariance functions (Damianou, Titsias, and Lawrence 2016) or through sampling (Damianou 2015; Salimbeni and Deisenroth 2017).

Variational approximations aren’t the only approach to approximate inference. The original work on deep Gaussian processes made use of MAP approximations (Lawrence and Moore 2007), which couldn’t propagate the uncertainty through the model at the data points but sustain uncertainty elsewhere. Since the variational approximation was proposed researchers have also considered sampling approaches (Havasi, Hernández-Lobato, and Murillo-Fuentes 2018) and expectation propagation (Bui et al. 2016).

Figure: Even the latest work on Bayesian neural networks has severe problems handling uncertainty. In this example, (Izmailov et al. 2019), methods even fail to interpolate through the data correctly or provide well calibrated error bars in regions where data is observed.

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

[edit]

Figure: Composition of linear functions just leads to a new linear function. Here you see the result of multiple affine transformations applied to a square in two dimensions.

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

[edit]

Figure: 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 further 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.

Stacked GPs (video by David Duvenaud)

[edit]

Figure: Visualization of mapping of a two dimensional space through a deep Gaussian process.

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

urllib.request.urlretrieve('https://raw.githubusercontent.com/lawrennd/talks/gh-pages/deepgp_tutorial.py','deepgp_tutorial.py')

# Late bind setup methods to DeepGP object
from deepgp_tutorial import initialize
from deepgp_tutorial import staged_optimize
from deepgp_tutorial import posterior_sample
from deepgp_tutorial import visualize
from deepgp_tutorial import visualize_pinball

import deepgp
deepgp.DeepGP.initialize=initialize
deepgp.DeepGP.staged_optimize=staged_optimize
deepgp.DeepGP.posterior_sample=posterior_sample
deepgp.DeepGP.visualize=visualize
deepgp.DeepGP.visualize_pinball=visualize_pinball

Olympic Marathon Data

[edit]

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 http://bit.ly/16kMKHQ

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())

Figure: Olympic marathon pace times since 1892.

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

[edit]

Figure: 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,
$$ \text{data} + \text{model} \stackrel{\text{compute}}{\rightarrow} \text{prediction} $$
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)
yt_sd=np.sqrt(yt_var)

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

Figure: Gaussian process fit to the Olympic Marathon data. The error bars are too large, perhaps due to the outlier from 1904.

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 make 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

[edit]

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'], 
                  kernels=[GPy.kern.RBF(hidden,ARD=True),
                           GPy.kern.RBF(x.shape[1],ARD=True)], # the kernels for each layer
                  num_inducing=50, back_constraint=False)

import deepgp

# Call the initalization
m.initialize()

Now optimize the model.

for layer in m.layers:
    layer.likelihood.variance.constrain_positive(warning=False)
m.optimize(messages=True,max_iters=10000)

m.staged_optimize(messages=(True,True,True))

Olympic Marathon Data Deep GP

Figure: Deep GP fit to the Olympic marathon data. Error bars now change as the prediction evolves.

Olympic Marathon Data Deep GP

Figure: Point samples run through the deep Gaussian process show the distribution of output locations.

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.

Figure: The mapping from input to the latent layer is broadly, with some flattening as time goes on. Variance is high across the input range.

Figure: The mapping from the latent layer to the output layer.

Olympic Marathon Pinball Plot

Figure: A pinball plot shows the movement of the ‘ball’ as it passes through each layer of the Gaussian processes. Mean directions of movement are shown by lines. Shading gives one standard deviation of movement position. At each layer, the uncertainty is reset. The overal uncertainty is the cumulative uncertainty from all the layers. There is some grouping of later points towards the right in the first layer, which also injects a large amount of uncertainty. Due to flattening of the curve in the second layer towards the right the uncertainty is reduced in final output.

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.

Gene Expression Example

[edit]

We now consider an example in gene expression. Gene expression is the measurement of mRNA levels expressed in cells. These mRNA levels show which genes are ‘switched on’ and producing data. In the example we will use a Gaussian process to determine whether a given gene is active, or we are merely observing a noise response.

Della Gatta Gene Data

[edit]

• 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())

Figure: Gene expression levels over time for a gene from data provided by Della Gatta et al. (2008). We would like to understand whethere there is signal in the data, or we are only observing noise.

• Want to detect if a gene is expressed or not, fit a GP to each gene Kalaitzis and Lawrence (2011).

Figure: The example is taken from the paper “A Simple Approach to Ranking Differentially Expressed Gene Expression Time Courses through Gaussian Process Regression.” 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.kern.lengthscale=50
_ = 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)
yt_sd=np.sqrt(yt_var)

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

Figure: Result of the fit of the Gaussian process model with the time scale parameter initialized to 50 minutes.

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

m_full2 = GPy.models.GPRegression(x,yhat)
m_full2.kern.lengthscale=2000
_ = m_full2.optimize() # Optimize parameters of covariance function

Figure: Result of the fit of the Gaussian process model with the time scale parameter initialized to 2000 minutes.

Now we try a model initialized with a lower noise.

m_full3 = GPy.models.GPRegression(x,yhat)
m_full3.kern.lengthscale=20
m_full3.likelihood.variance=0.001
_ = m_full3.optimize() # Optimize parameters of covariance function

Figure: Result of the fit of the Gaussian process model with the noise initialized low (standard deviation 0.1) and the time scale parameter initialized to 20 minutes.

Figure:

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, 
                  inits=inits, 
                  kernels=kernels, # the kernels for each layer
                  num_inducing=20, back_constraint=False)

m.initialize()
m.staged_optimize()

Della Gatta Gene Data Deep GP

[edit]

Figure: Deep Gaussian process fit to the Della Gatta gene expression data.

Della Gatta Gene Data Deep GP

Figure: Deep Gaussian process samples fitted to the Della Gatta gene expression data.

Della Gatta Gene Data Latent 1

Figure: Gaussian process mapping from input to latent layer for the della Gatta gene expression data.

Della Gatta Gene Data Latent 2

Figure: Gaussian process mapping from latent to output layer for the della Gatta gene expression data.

TP53 Gene Pinball Plot

Figure: A pinball plot shows the movement of the ‘ball’ as it passes through each layer of the Gaussian processes. Mean directions of movement are shown by lines. Shading gives one standard deviation of movement position. At each layer, the uncertainty is reset. The overal uncertainty is the cumulative uncertainty from all the layers. Pinball plot of the della Gatta gene expression data.

Step Function

[edit]

Next we consider a simple step function data set.

num_low=25
num_high=25
gap = -.1
noise=0.0001
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

Figure: Simulation study of step function data artificially generated. Here there is a small overlap between the two lines.

Step Function Data GP

We can fit a Gaussian process to the step function data using GPy as follows.

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

Where GPy.models.GPRegression() gives us a standard GP regression model with exponentiated quadratic covariance function.

The model is optimized using m_full.optimize() which calls an L-BGFS gradient based solver in python.

Figure: Gaussian process fit to the step function data. Note the large error bars and the over-smoothing of the discontinuity. Error bars are shown at two standard deviations.

The resulting fit to the step function data shows some challenges. In particular, the over smoothing at the discontinuity. If we know how many discontinuities there are, we can parameterize them in the step function. But by doing this, we form a semi-parametric model. The parameters indicate how many discontinuities are, and where they are. They can be optimized as part of the model fit. But if new, unforeseen, discontinuities arise when the model is being deployed in practice, these won’t be accounted for in the predictions.

Step Function Data Deep GP

[edit]

First we initialize a deep Gaussian process with three latent layers (four layers total). Within each layer we create a GP with an exponentiated quadratic covariance (GPy.kern.RBF).

At each layer we use 20 inducing points for the variational approximation.

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, 
                  inits=inits, 
                  kernels=kernels, # the kernels for each layer
                  num_inducing=20, back_constraint=False)

Once the model is constructed we initialize the parameters, and perform the staged optimization which starts by optimizing variational parameters with a low noise and proceeds to optimize the whole model.

m.initialize()
m.staged_optimize()

We plot the output of the deep Gaussian process fitted to the stpe data as follows.

The deep Gaussian process does a much better job of fitting the data. It handles the discontinuity easily, and error bars drop to smaller values in the regions of data.

Figure: Deep Gaussian process fit to the step function data.

Step Function Data Deep GP

The samples of the model can be plotted with the helper function from teaching_plots.py, model_sample

The samples from the model show that the error bars, which are informative for Gaussian outputs, are less informative for this model. They make clear that the data points lie, in output mainly at 0 or 1, or occasionally in between.

Figure: Samples from the deep Gaussian process model for the step function fit.

The visualize code allows us to inspect the intermediate layers in the deep GP model to understand how it has reconstructed the step function.

Figure: From top to bottom, the Gaussian process mapping function that makes up each layer of the resulting deep Gaussian process.

A pinball plot can be created for the resulting model to understand how the input is being translated to the output across the different layers.

Figure: Pinball plot of the deep GP fitted to the step function data. Each layer of the model pushes the ‘ball’ towards the left or right, saturating at 1 and 0. This causes the final density to be be peaked at 0 and 1. Transitions occur driven by the uncertainty of the mapping in each layer.

import pods

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

Motorcycle Helmet Data

[edit]

Figure: Motorcycle helmet data. The data consists of acceleration readings on a motorcycle helmet undergoing a collision. The data exhibits heteroschedastic (time varying) noise levles and non-stationarity.

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

Motorcycle Helmet Data GP

[edit]

Figure: Gaussian process fit to the motorcycle helmet accelerometer data.

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, 
                  inits=inits, 
                  kernels=kernels, # the kernels for each layer
                  num_inducing=20, back_constraint=False)



m.initialize()

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

Motorcycle Helmet Data Deep GP

[edit]

Figure: Deep Gaussian process fit to the motorcycle helmet accelerometer data.

Motorcycle Helmet Data Deep GP

Figure: Samples from the deep Gaussian process as fitted to the motorcycle helmet accelerometer data.

Motorcycle Helmet Data Latent 1

Figure: Mappings from the input to the latent layer for the motorcycle helmet accelerometer data.

Motorcycle Helmet Data Latent 2

Figure: Mappings from the latent layer to the output layer for the motorcycle helmet accelerometer data.

Motorcycle Helmet Pinball Plot

Figure: Pinball plot for the mapping from input to output layer for the motorcycle helmet accelerometer data.

Robot Wireless Data

[edit]

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.

data=pods.datasets.robot_wireless()

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

Figure: Ground truth movement for the position taken while recording the multivariate time-course of wireless access point signal strengths.

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

Figure: Output dimension 1 from the robot wireless data. This plot shows signal strength changing over time.

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

Figure: Gaussian process fit to the Robot Wireless dimension 1.

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, 
                  kernels=kernels,
                  num_inducing=50, back_constraint=False)
m.initialize()

m.staged_optimize(messages=(True,True,True))

Robot WiFi Data Deep GP

[edit]

Figure: Fit of the deep Gaussian process to dimension 1 of the robot wireless data.

Robot WiFi Data Deep GP

Figure: Samples from the deep Gaussian process fit to dimension 1 of the robot wireless data.

Robot WiFi Data Latent Space

}

Figure: Inferred two dimensional latent space of the model for the robot wireless data.

‘High Five’ Motion Capture Data

[edit]

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

Shared LVM

[edit]

Figure: Shared latent variable model structure. Here two related data sets are brought together with a set of latent variables that are partially shared and partially specific to one of the data sets.

Figure: Latent spaces of the ‘high five’ data. The structure of the model is automatically learnt. One of the latent spaces is coordinating how the two figures walk together, the other latent spaces contain latent variables that are specific to each of the figures separately.

Subsample of the MNIST Data

[edit]

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_openml

mnist = fetch_mldata('MNIST original')

Sub-sample the dataset to make the training faster.

import numpy as np

np.random.seed(0)
digits = [0,1,2,3,4]
N_per_digit = 100
Y = []
labels = []
for d in digits:
    imgs = mnist['data'][mnist['target']==d]
    Y.append(imgs[np.random.permutation(imgs.shape[0])][:N_per_digit])
    labels.append(np.ones(N_per_digit)*d)
Y = np.vstack(Y).astype(np.float64)
labels = np.hstack(labels)
Y /= 255.

Fitting a Deep GP to a the MNIST Digits Subsample

[edit]

Thanks to: Zhenwen Dai and Neil D. Lawrence

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

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],
                  Y,
                  kernels=[GPy.kern.RBF(num_hidden_2,ARD=True), 
                           GPy.kern.RBF(num_latent,ARD=False)], 
                  num_inducing=50, back_constraint=False, 
                  encoder_dims=[[200],[200]])

Initialization

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:
    layer.kern.variance.fix(warning=False)
    layer.likelihood.variance.fix(warning=False)

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

m.optimize(messages=False,max_iters=100)

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:
    layer.kern.variance.constrain_positive(warning=False)
m.optimize(messages=False,max_iters=100)

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

for layer in m.layers:
    layer.likelihood.variance.constrain_positive(warning=False)
m.optimize(messages=True,max_iters=10000)

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.

Figure: Latent space for the deep Gaussian process learned through unsupervised learning and fitted to a subset of the MNIST digits subsample.

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

m.obslayer.kern.lengthscale

Figure: Visualisation of the intermediate layer, plot of dimension 1 vs dimension 0.

Figure: Visualisation of the intermediate layer, plot of dimension 1 vs dimension 0.

Figure: Visualisation of the intermediate layer, plot of dimension 1 vs dimension 0.

Figure: Visualisation of the intermediate layer, plot of dimension 1 vs dimension 0.

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

Figure: These digits are produced by taking a tour of the two dimensional latent space (as described by a Gaussian process sample) and mapping the tour into the data space. We visualize the mean of the mapping in the images.

Deep Health

[edit]

Figure: The deep health model uses different layers of abstraction in the deep Gaussian process to represent information about diagnostics and treatment to model interelationships between a patients different data modalities.

From a machine learning perspective, we’d like to be able to interrelate all the different modalities that are informative about the state of the disease. For deep health, the notion is that the state of the disease is appearing at the more abstract levels, as we descend the model, we express relationships between the more abstract concept, that sits within the physician’s mind, and the data we can measure.

Thanks!

For more information on these subjects and more you might want to check the following resources.

• twitter: @lawrennd
• podcast: The Talking Machines
• newspaper: Guardian Profile Page
• blog: http://inverseprobability.com

References