A deep learning first explanation by Joost van Amersfoort
Until 2013, Gaussian Processes (GPs) were roughly as popular as deep learning, according to Google Trends. Since then deep learning has seen a 30x increase in interest, while interest in GPs remained unchanged (although the UK is 3x more interested relative to the US!). Despite the community not increasing as fast, there's been steady research happening with GPs. In this blog, I'll explain one of the main research advances in GPs from a deep learning first perspective by deriving it similarly to the Variational Auto-Encoder (VAE)!
GPs come out-of-the-box with uncertainty, some interpretability, and rigorous mathematical foundations. So it's not surprising that they remain a topic of great interest to research! However compared to deep learning, GPs are more difficult to scale, both in terms of training data set size, and the dimensionality of the data itself. On both fronts the GP community is making progress, but for this blog I'll focus on one of the proposed solutions to scaling to large data set sizes: variational inference in GPs. I'll use notation from the deep learning literature and attempt to rederive a famous result in GPs using steps that resemble deriving the VAE objective. I'll also include as much intuition as possible and by doing this I hope to make the GP literature more accessible for those with only experience in deep learning.
For the purpose of this blog, I assume that you have some familiarity with the Variational Auto-Encoder and variational inference in general. A good reference to get up to speed with both those concepts is this blog by Jaan Altosaar, specifically the section titled "The probability model perspective". Familiarity with GPs is not necessary, but if you would like a gentle explanation with many beautiful visualisations then the Distill article on GPs is great.
Let's start by recalling the assumptions that underlie the VAE. Our first assumption is that there is a random variable from which it is possible to generate our data set . Inversely, if we have a data point , we also want to be able find ("infer") its corresponding . To make things interesting, we restrict such that it is lower dimensional than our data points and easy to sample from.
It turns out that if we were given the and pairs, then it's straightforward to train a model to predict from , i.e. . However we only have our data set , and we need to figure out how to structure , such that it becomes usable, i.e. , and afterwards we can model . Unfortunately, computing the true value of is intractable (it requires evaluating all possible values of !). Instead we will try to find a good approximate distribution, which we call q. We decide that q will be a Gaussian distribution parametrised by , which lends itself well for applying the reparametrisation trick and also allows easy sampling of new data points (fun!).
The process of finding the best approximation q(z|x) to p(z|x) is called Variational Inference!
We measure the difference between and using the KL divergence and we are interested in finding the set of parameters that minimise it:
Rewriting the KL divergence we obtain:
Since is independent of , we can take it out of the integral and because is a probability distribution (which integrate to 1), we simply have left. Also it doesn't depend on , so we can disregard it for this minimisation. We now take the integral, express it as an expectation and flip the sign (turning the argmin into and argmax) and obtain the evidence lower bound (ELBO):
The last formula is one we can work with! is the distribution we are interested in and we will also model it using a neural network, but with parameters . We choose to be , which makes the KL with easy to compute. The only problem we have left is how to compute the expectation over !
To evaluate the gradient of the expectation, we would like to use samples of . However, sampling from is not a differentiable operation and won't allow us to backpropagate back to . The key insight from the reparametrisation trick (introduced concurrently by Kingma et al., Rezende et al., and Titsias et al.) is that we can instead sample from a standard Gaussian and transform its samples in a deterministic way such that are samples from q:
With and being outputs of the neural network . If we plug this into the previous equation, we obtain:
This equation is fully differentiable, so we can optimise it using backpropagation! A simple implementation of a model based on this equation can be found in the PyTorch example of the VAE. With this result in hand, we will go look at the situation in GPs.
We start by looking at the definition of what is sometimes called the "full" Gaussian process and continue with the more practical sparse Gaussian process.
A Gaussian process is defined by a prior mean and covariance function, respectively and , of the data set :
In this blog, we assume our GP has zero prior mean and replace by . This is not a limitation of GPs, but it makes the notation easier. The covariance function, also referred to as kernel, is where we can apply any knowledge we have of the data we are modelling. A common choice is the squared exponential kernel:
which has two hyper parameters: the length scale and the output variance . The length scale represents an assumption of the amount of variation in your inputs, the higher the value the smoother the underlying function is assumed to be (and the more we can confidently extrapolate). The output variance is a scale factor and depends on the domain of the input. The core assumption in the squared exponential kernel is that data points that are close in input ("similar"), have high covariance ("similar" labels). There are many other options, such as linear kernels, periodic kernels and combinations are also possible. See David Duvenaud's excellent kernel cookbook for more.
In GPs, we further assume that we can only observe noisy labels (with noise ) of data. We define as the function values at : :
We are interested in finding the predictive distribution for new data point , because it will allow us to make prediction for unseen data.
The joint distribution of the noisy labels and unseen data is:
With a vector of kernel distances between the test data point and the training data. For more detail refer to equation 2.21 of the GPML book. If we condition on (the observed data), we obtain the equation for the mean and covariance of the predictive distribution:
The equation for contains an inverse of the covariance matrix of the data set. We can only stably invert covariance matrices up to a few thousand points. After that it becomes very slow, even on modern GPUs, and can also fail due to numeric problems. One way to go around this limitation is to exploit the fact that there is often a lot of redundancy in the training set. By using only the points that are most informative we can drastically reduce computation. GPs that work on such a reduced data set are called "Sparse Gaussian Processes".
The first Sparse GPs relied on smartly picking a subset of the training data, this subset is also referred to as the "inducing points", and computing the inverse on those. This allows us to reduce the computation from depending on the number of data points to depending on the number of induced points. See Quiñonero-Candela et al. for a nice overview.
Later methods removed the reliance on selecting an actual subset of the data points (Snelson et al.) by introducing the concept of "pseudo" inducing points. It was Titsias who introduced the first variational approximation to optimising sparse GPs, where the pseudo inducing points are optimised as variational parameters. In contrast to previous approaches, the "Titsias approximation" made optimising the pseudo inducing points part of the variational objective, providing a unified and rigorous way to optimise them.
The definition of sparse GPs is a bit different from full GPs. Below we follow the derivation of Hensman et al.:
With containing values of the function at the (pseudo-) inducing points and . Notice that the matrix inverse is on now, making it dependent on the number of inducing points we pick. It no longer grows with the data set. The difference between this approximation and the full GP is given by the following KL divergence:
Note that we write and not as second distribution, because may not fully overlap with Y if we use pseudo inducing points. If contains all training function values, meaning that , then and so . In that case the KL divergence is zero, because both distributions are degenerate and we recover the full GP. However we lost the computational gain, because the matrix inverse is of the order of the original data set again.
We introduce an approximate posterior distribution and attempt to minimise its distance to the true posterior . We start from a similar divergence as in the VAE and show we can find a bound we can directly optimise with respect to the pseudo inducing points.
The KL we are interested in minimising is:
Intuitively, we can interpret this as the difference between having only access to the inducing points as compared to having access to both the inducing and the original data points. If this difference is small, then q holds all the information we need.
Like before, we will write out the KL divergence in (now) familiar components:
The first to the second step is done using Bayes' rule. The second to the third step is writing out the fraction inside the log into separate components. The last step rewrites the integral into familiar components (KL divergence, likelihood and marginal likelihood). The marginal log-likelihood is over instead of as we saw in the VAE. To evaluate the likelihood, we use the fact that given the inducing points, the function values factorise:
The approximate posterior is itself a Gaussian Process, with a mean and variance function that depend on the inducing variables. Following this blog by James Hensman, we pick a particular mean and variance function for q (which we are free to do!), such that the ELBO is computable:
This step might seem abstract, but it's common in variational inference to pick such that the ELBO becomes tractable. The more flexible we allow, the better it can approximate the true posterior. , , and are all variational parameters that can be optimised through gradient descent. Interestingly, this means that the inducing points themselves, , have become part of the objective. This is different way of thinking than in the VAE, where we only care about updating the parameters of the deep network. For a detailed derivation, see also this tutorial by Yarin Gal and Mark van der Wilk.
In this blog I have shown that performing variational inference in VAEs and GPs is very similar. In both cases we start out with an intractable equation and introduce a new distribution to create a tractable approximation. We try to find the best approximation to the original problem (or alternatively minimise the KL divergence!).
If you found this interesting and are curious to find out more about GPs, then there is a considerable amount of literature available. I can recommend the deep GP paper and variational extensions of it.
Lastly I would like to bring your attention to my own work, which combines the benefits of GPs with the flexibility of deep learning: vDUQ. vDUQ is an instance of Deep Kernel Learning and uses the inducing point approximation we discussed above! Interestingly, the inducing points are placed in the feature space of a deep neural network, allowing only a handful of inducing points to be used for complicated datasets such as CIFAR-10.