Gaussian Process Classification

Gaussian Processes provide an elegant framework for regression, offering not just predictions but principled uncertainty quantification through closed-form posterior distributions. But what happens when our target variable is no longer continuous? What if we need to classify emails as spam or not-spam, diagnose patients as healthy or diseased, or predict whether a financial transaction is fraudulent?

The fundamental obstacle: Classification outputs are discrete—typically binary labels in {0, 1} or class probabilities in [0, 1]. The Gaussian likelihood that made GP regression so tractable is fundamentally incompatible with these outputs. We cannot assume that class labels are generated by adding Gaussian noise to a latent function. This incompatibility shatters the beautiful closed-form solutions we derived for regression and forces us into the realm of approximate inference.

Let's first recall why GP regression works so elegantly, then identify exactly where classification breaks this framework.

GP Regression Recap:

In GP regression, we model observations as:

$$y_i = f(\mathbf{x}_i) + \epsilon_i, \quad \epsilon_i \sim \mathcal{N}(0, \sigma_n^2)$$

where $f \sim \mathcal{GP}(m, k)$ is our latent function. The key property is that:

$$p(y_i | f_i) = \mathcal{N}(y_i | f_i, \sigma_n^2)$$

This Gaussian likelihood, combined with the Gaussian process prior, yields a Gaussian posterior over the latent function. The product of two Gaussians is Gaussian—this is the conjugacy that enables closed-form inference.

The Classification Setup:

For binary classification with labels $y \in {0, 1}$, we need:

$$p(y_i = 1 | f_i) = \sigma(f_i)$$

where $\sigma: \mathbb{R} \to [0, 1]$ is a squashing function (also called a response function or inverse link function) that maps the unbounded latent function value to a valid probability.

The Latent Function Interpretation:

We can think of classification as a two-stage process:

1. Latent function evaluation: Draw $f_i = f(\mathbf{x}_i)$ from the GP
2. Probabilistic classification: Map $f_i$ through squashing function $\sigma$ to get class probability

The latent function $f$ represents an underlying 'propensity' or 'log-odds' toward the positive class. Large positive values of $f$ indicate high probability of class 1; large negative values indicate high probability of class 0.

This interpretation is powerful: we maintain the GP's modeling of correlations between inputs through the kernel, while the squashing function handles the mapping to valid probabilities.

The most common choice for binary classification is the logistic sigmoid function:

$$\sigma(f) = \frac{1}{1 + e^{-f}} = \frac{e^f}{1 + e^f}$$

This yields the Bernoulli likelihood:

$$p(y | f) = \sigma(f)^y \cdot (1 - \sigma(f))^{1-y} = \text{Bernoulli}(y | \sigma(f))$$

Or equivalently, writing it as a single expression:

$$p(y | f) = \sigma(yf) \quad \text{where } y \in {-1, +1}$$

(Note: The second form uses labels in ${-1, +1}$ rather than ${0, 1}$, which simplifies many derivations.)

Why Logistic Sigmoid?

The logistic function arises naturally from several perspectives:

1. Maximum entropy: Among all distributions over ${0, 1}$ with given mean, the Bernoulli-logistic model maximizes entropy.

2. Exponential family: The logistic function is the canonical link for the Bernoulli exponential family.

3. Log-odds interpretation: The latent function $f$ directly represents the log-odds $\log(p/(1-p))$, making the model interpretable.

4. Asymptotic behavior: Smooth saturation at extremes provides numerical stability.

The Full Likelihood:

For a dataset $\mathcal{D} = {(\mathbf{x}i, y_i)}{i=1}^n$ and latent function values $\mathbf{f} = [f_1, ..., f_n]^\top$:

$$p(\mathbf{y} | \mathbf{f}) = \prod_{i=1}^n p(y_i | f_i) = \prod_{i=1}^n \sigma(f_i)^{y_i}(1-\sigma(f_i))^{1-y_i}$$

This factorizes across data points because, conditioned on $f$, observations are independent.

An alternative to the logistic sigmoid is the probit function—the cumulative distribution function (CDF) of the standard normal distribution:

$$\Phi(f) = \int_{-\infty}^f \mathcal{N}(z | 0, 1) , dz = \frac{1}{2}\left[1 + \text{erf}\left(\frac{f}{\sqrt{2}}\right)\right]$$

The probit likelihood is:

$$p(y | f) = \Phi(f)^y \cdot (1 - \Phi(f))^{1-y}$$

Latent Variable Interpretation:

The probit model has an elegant latent variable interpretation. Consider:

$$y = \mathbf{1}[f + \epsilon > 0], \quad \epsilon \sim \mathcal{N}(0, 1)$$

Then: $$p(y = 1 | f) = p(f + \epsilon > 0) = p(\epsilon > -f) = 1 - \Phi(-f) = \Phi(f)$$

This interpretation is particularly useful for multi-class extensions and connects to classical statistical models.

Comparing Logistic and Probit:

Both functions are monotonically increasing, symmetric around 0, and map $\mathbb{R} \to (0, 1)$. They are nearly indistinguishable for moderate values of $f$ (approximately $\sigma(f) \approx \Phi(1.7 \cdot f)$).

The key difference is in the tails: the logistic function has heavier tails (slower decay to 0 and 1), making it slightly more robust to outliers. The probit function's lighter tails mean that extreme observations have more influence on the latent function.

Practical Guidance:

• Use logistic for most applications (computationally simpler, widely understood)
• Use probit when you need the latent variable interpretation, or for theoretical analysis
• The choice rarely matters in practice—focus on the kernel design and inference method instead

Now we arrive at the central challenge of GP classification. By Bayes' theorem, the posterior over latent function values is:

$$p(\mathbf{f} | \mathbf{y}, X) = \frac{p(\mathbf{y} | \mathbf{f}) \cdot p(\mathbf{f} | X)}{p(\mathbf{y} | X)}$$

where:

• $p(\mathbf{f} | X) = \mathcal{N}(\mathbf{f} | \mathbf{m}, K)$ is the GP prior (Gaussian)
• $p(\mathbf{y} | \mathbf{f}) = \prod_i \sigma(f_i)^{y_i}(1-\sigma(f_i))^{1-y_i}$ is the likelihood (non-Gaussian)
• $p(\mathbf{y} | X) = \int p(\mathbf{y} | \mathbf{f}) p(\mathbf{f} | X) d\mathbf{f}$ is the marginal likelihood (intractable)

Visualizing the Problem:

For a single latent variable $f$, consider:

$$p(f | y, x) \propto \underbrace{\sigma(f)^y (1-\sigma(f))^{1-y}}{\text{Bernoulli likelihood}} \cdot \underbrace{\mathcal{N}(f | m, k)}{\text{GP prior}}$$

The prior is a Gaussian (bell curve). The likelihood is a sigmoid (S-curve) or its complement. Their product is skewed and asymmetric—manifestly non-Gaussian.

For $n$ data points, we face an $n$-dimensional integral. Even for modest $n$, direct numerical integration becomes computationally infeasible.

The Two Intractable Computations:

1. Posterior computation: Finding $p(\mathbf{f} | \mathbf{y}, X)$ for the training points
2. Predictive distribution: Finding $p(f_* | \mathbf{y}, X)$ for new test points

Both require integrating out the latent function values, which cannot be done analytically.

Even if we could compute the posterior $p(\mathbf{f} | \mathbf{y}, X)$, making predictions at a new test point $\mathbf{x}_*$ involves additional integration.

Step 1: Latent Predictive Distribution

First, we need the distribution of the latent function at the test point:

$$p(f_* | \mathbf{y}, X, \mathbf{x}*) = \int p(f* | \mathbf{f}, X, \mathbf{x}_*) \cdot p(\mathbf{f} | \mathbf{y}, X) , d\mathbf{f}$$

The first term is the GP predictive given training latents—this is Gaussian: $$p(f_* | \mathbf{f}, X, \mathbf{x}*) = \mathcal{N}(f* | \mu_, \sigma_^2)$$

with: $$\mu_* = m(\mathbf{x}*) + \mathbf{k}^\top K^{-1}(\mathbf{f} - \mathbf{m})$$ $$\sigma_^2 = k(\mathbf{x}*, \mathbf{x}) - \mathbf{k}_^\top K^{-1} \mathbf{k}_*$$

But averaging this Gaussian over the non-Gaussian posterior $p(\mathbf{f} | \mathbf{y}, X)$ has no closed form.

Step 2: Class Probability Prediction

Once we have (or approximate) the latent predictive distribution $p(f_* | \mathbf{y}, X, \mathbf{x}_*)$, the class probability is:

$$p(y_* = 1 | \mathbf{y}, X, \mathbf{x}*) = \int \sigma(f) \cdot p(f_ | \mathbf{y}, X, \mathbf{x}*) , df*$$

This is a one-dimensional integral, which is more tractable—but still requires knowledge of the latent predictive distribution.

Special Case: Probit Likelihood with Gaussian Approximation

If we approximate $p(f_* | \mathbf{y}, X, \mathbf{x}*) \approx \mathcal{N}(\mu, \sigma_^2)$, and use the probit likelihood:

$$\int \Phi(f_) \cdot \mathcal{N}(f_ | \mu_, \sigma_^2) , df_* = \Phi\left(\frac{\mu_}{\sqrt{1 + \sigma_^2}}\right)$$

This closed-form solution for probit is one reason it's sometimes preferred for theoretical analysis. For logistic, this integral must be approximated numerically.

Since exact inference is impossible, the GP classification literature has developed several approximation strategies. Each makes different trade-offs between accuracy, computational cost, and implementation complexity.

The Core Approaches:

All methods approximate the intractable posterior $p(\mathbf{f} | \mathbf{y}, X)$ with a tractable distribution $q(\mathbf{f})$. The differences lie in how $q$ is chosen and optimized.

Why Gaussian Approximations?

Almost all practical GP classification methods approximate the posterior with a Gaussian:

$$q(\mathbf{f}) = \mathcal{N}(\mathbf{f} | \boldsymbol{\mu}, \Sigma)$$

This choice is not arbitrary:

1. Computational tractability: Gaussian distributions are closed under marginalization, conditioning, and linear transformation. This makes predictions tractable.

2. Natural extension of regression: The true posterior in GP regression is Gaussian. Using Gaussian approximations for classification maintains algorithmic similarity.

3. Parameterization: A Gaussian over $n$ training points requires $O(n)$ mean parameters and $O(n^2)$ covariance parameters—expensive, but feasible.

4. Uncertainty propagation: Gaussian approximations naturally propagate uncertainty through the prediction pipeline.

The Question of Quality:

The different methods (Laplace, EP, VI) correspond to different ways of fitting the parameters $(\boldsymbol{\mu}, \Sigma)$. They differ in:

• How well they capture the true posterior's location and spread
• Computational cost per iteration
• Convergence properties
• Quality of marginal likelihood approximation (for hyperparameter learning)

To unify our treatment of GP classification methods, we establish the common mathematical framework that all approaches share.

The Model:

$$\mathbf{f} | X \sim \mathcal{N}(\mathbf{m}, K) \quad \text{(GP prior)}$$ $$y_i | f_i \stackrel{\text{ind}}{\sim} \text{Bernoulli}(\sigma(f_i)) \quad \text{(conditional independence)}$$

The Inference Goals:

1. Approximate the posterior: Find $q(\mathbf{f}) \approx p(\mathbf{f} | \mathbf{y}, X)$

2. Approximate the marginal likelihood: Estimate $p(\mathbf{y} | X) = \int p(\mathbf{y} | \mathbf{f}) p(\mathbf{f} | X) d\mathbf{f}$ for hyperparameter optimization

3. Make predictions: Compute $p(y_* = 1 | \mathbf{y}, X, \mathbf{x}_*)$ for new inputs

Key Quantities:

Define the log-likelihood and its derivatives:

$$\psi_i(f_i) = \log p(y_i | f_i)$$ $$ abla \psi_i = \frac{\partial \psi_i}{\partial f_i}$$ $$ abla^2 \psi_i = \frac{\partial^2 \psi_i}{\partial f_i^2}$$

For the Logistic Likelihood:

$$\psi_i = y_i \log \sigma(f_i) + (1-y_i) \log(1-\sigma(f_i))$$

Using the identity $\sigma'(f) = \sigma(f)(1-\sigma(f))$:

$$ abla \psi_i = y_i - \sigma(f_i) = y_i - \pi_i$$

$$ abla^2 \psi_i = -\sigma(f_i)(1-\sigma(f_i)) = -\pi_i(1-\pi_i)$$

where $\pi_i = \sigma(f_i)$ is the predicted probability.

Key Observations:

1. The gradient $ abla \psi_i$ is the residual between the label and predicted probability
2. The Hessian $ abla^2 \psi_i$ is always negative (log-likelihood is concave)
3. The Hessian magnitude is maximized when $\pi_i = 0.5$ and minimized at extremes

These properties ensure the log-posterior is concave (since the log of a Gaussian is quadratic), which means the Laplace approximation finds a unique global maximum.

We have established the fundamental challenge that defines all of GP classification: the incompatibility between Gaussian process priors and classification likelihoods.

What's Next:

With the problem clearly defined, we now turn to solutions. The next page introduces the Laplace approximation—the simplest and most widely-used method for GP classification. We'll derive the algorithm, understand its geometric interpretation, and learn when it works well and when more sophisticated methods are needed.