Home » Understanding the Fundamentals of Approximate Inference
A topography map

Understanding the Fundamentals of Approximate Inference

In the realm of machine learning, inference plays a crucial role in enabling algorithms to make predictions based on data. However, many real-world datasets are complex and often intractable, requiring a specialized form of inference known as approximate inference. This article will explore the fundamentals of approximate inference, including its relationship to optimization, key algorithms such as Expectation Maximization, MAP Inference, and Variational Inference, as well as the concept of learned approximate inference.

Inference as Optimization

Inference can be thought of as an optimization problem, where the goal is to find the parameters that best explain the observed data. In a probabilistic model, the parameters define the likelihood of the data given the model. The inference process involves finding the parameters that maximize the likelihood of the data given the model, also known as maximum likelihood estimation.

However, when dealing with complex and large datasets, the likelihood function can be difficult to optimize directly. Approximate inference provides a way to bypass this issue by approximating the intractable likelihood function with a tractable approximation, often in the form of a simpler probabilistic model. The goal of approximate inference is to find the parameters of the approximation that best match the parameters of the original model.

Expectation Maximization

Gradients are all around us

Expectation Maximization (EM) is a popular algorithm used in approximate inference. It is an iterative algorithm that alternates between two steps: the expectation (E) step, where the algorithm estimates the expected value of the hidden variables given the current estimates of the parameters, and the maximization (M) step, where the algorithm updates the parameters based on the expected values of the hidden variables.

The algorithm continues to alternate between the E and M steps until convergence, at which point the parameters are said to be maximized. EM is particularly useful when the original likelihood function has hidden variables, which can be difficult to optimize directly.

MAP Inference and Sparse Coding

Maximum a Posteriori (MAP) inference is another popular algorithm used in approximate inference. It involves finding the parameters that maximize the posterior distribution of the model given the data. Unlike maximum likelihood estimation, which only considers the likelihood of the data given the model, MAP inference also incorporates prior knowledge about the parameters through the use of a prior distribution.

Sparse coding is a specific application of MAP inference where the goal is to find a sparse representation of the data in terms of a set of basis functions. In this context, the sparse representation is achieved by using a Laplacian prior on the parameters, which encourages sparsity by penalizing large values.

Variational Inference and Learning

Variational Inference is a form of approximate inference that involves finding the parameters of a tractable approximation to the intractable posterior distribution. The approximation is found by minimizing the KL divergence between the approximation and the true posterior distribution, which measures the difference between the two distributions.

An expanse of Gaussian gradients

Variational Inference has become increasingly popular in recent years due to its ability to scale to large datasets and its connection to deep learning, where it is often used as a building block for variational autoencoders and other deep generative models.

Learned Approximate Inference

Learned approximate inference is a new area of research that focuses on using machine learning techniques to learn the approximation for the intractable posterior distribution. The goal is to automatically learn a tractable approximation that is optimized for the specific problem at hand, rather than relying on hand-designed approximations used in traditional approximate inference algorithms.

This field is still in its infancy, but it has the potential to revolutionize the way approximate inference is performed, as it can lead to more accurate and efficient solutions for complex datasets. For example, instead of using a Gaussian approximation as in Variational Inference, a neural network could be trained to approximate the posterior distribution. This would allow for more flexible and sophisticated approximations, tailored to the specific problem and data at hand.

One promising approach to learned approximate inference is the use of normalizing flow models. Normalizing flows are a class of generative models that transform a simple random variable into a complex distribution. By using normalizing flows as the approximation, the parameters can be optimized end-to-end using gradient-based methods.

Another approach is the use of reinforcement learning, where the approximate inference algorithm is treated as an agent and trained using a reward function that measures the accuracy of the approximation. This approach has shown promising results in a number of experimental studies and has the potential to lead to significant advances in the field of approximate inference.

In conclusion, approximate inference is an important area of machine learning that enables algorithms to make predictions based on complex and intractable datasets. The field of approximate inference has a rich history, including popular algorithms such as EM, MAP inference, and Variational Inference, as well as the emerging area of learned approximate inference. As the field continues to evolve, it is likely that we will see even more sophisticated and effective solutions for complex inference problems.


  1. Blei, D. M., Kucukelbir, A., & McAuliffe, J. D. (2017). Variational inference: A review for statisticians. Annual Review of Statistics and Its Application, 4(1), 355-378.
  2. Kingma, D. P., & Ba, J. (2014). Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980.
  3. Rezende, D. J., & Mohamed, S. (2015). Variational inference with normalizing flows. arXiv preprint arXiv:1505.05770.
  4. Wainwright, M. J. (2008). Graphical models, exponential families, and variational inference. Foundation and Trends in Machine Learning, 1(1-2), 1-305.

More Reading

Post navigation

Leave a Comment

Leave a Reply