Tl;dr Approximate Inference methods made easy

<p>A closed-form solution to a machine learning model is one that can be written down on a sheet of paper using a finite number of standard mathematical operations. For example, linear models have closed-form solutions IF the design covariance matrix is invertible, otherwise we obtain a solution using iterative optimisation.</p> <p><img alt="" src="https://miro.medium.com/v2/1*gOOz91wskRFGwRqOMTgUtA.png" style="width:700px" /></p> <p>Bayesian models do not typically have exact closed-form solutions for their posterior distributions. One thing that typically helps is choosing simple models, Gaussian likelihood functions and conjugate priors. A prior distribution is said to be&nbsp;<em>conjugate</em>&nbsp;to a likelihood function if the resulting posterior belongs to the same distribution family as the prior.</p> <p><img alt="" src="https://miro.medium.com/v2/resize:fit:700/1*xrCq2K9jbBp84qYU5UD5oQ.png" style="height:251px; width:700px" /></p> <p>Bayesian linear regression is a model that typically assumes Gaussian priors over both the regression coefficients and the likelihood function. When we update the prior with the observed data (using Bayes&rsquo; theorem), the resulting posterior distribution for the regression coefficients will also follow a normal distribution. This can be written down analytically and sampled using standard methods in Python.</p> <p><a href="https://chleon.medium.com/tl-dr-approximate-inference-methods-made-easy-c9652947bc04"><strong>Visit Now</strong></a></p>