Machine Learning: 5 Part Series About How Machine Learning Works

[stag_dropcap font_size="40px" style="squared"]T[/stag_dropcap]here has been alot of talk about Machine learning – most notably for digital marketers and SEOs in regards to recent Google algorithm updates. But have you ever asked yourself “what is machine learning”, or wondered how it actually works? BloomReach engineer Srinath Sridhar walks through probability, Bayesian models and machine learning in this 5 part video series.

Machine learning, a branch of artificial intelligence, is about the construction and study of systems that can learn from data.

Machine Learning (Part 1 of 5): Fundamentals


This lecture goes over some fundamental definitions of statistics that are needed for any rigorous analysis of machine learning algorithms. They will define random variable, sample space, probability, expectation, standard deviation and variance and go over examples of discrete and continuous probability distributions. They specifically spend time on uniform, binomial and normal (Gaussian) distributions. They briefly mention other distributions such as poisson, exponential, geometric and negative binomial but these are left to the you to understand them in detail.

Machine Learning (Part 2 of 5): Probability


This lecture continues to build on fundamental definitions of statistics. This time they talk about joint, conditional probabilities and conditional expectations. They go over an example of conditional probabilities in the real-world and talk about how they impact analysis. They conclude with talking about Bayes rule which will set you up to talk about a Bayesian classifier in the following lecture.

Machine Learning (Part 3 of 5): Naive Bayes Classifier


This lecture continues to build on Bayes rule that was taught last time. They define training and testing data sets and build a Bayesian classifier. Specifically they define prior, likelihood and posterior. They use titles and product descriptions from a retailer and attempt to find the top level category that the product is listed under. The likelihood corresponds to per category word frequencies and the prior correspond to the number of products under each category. They talk about running into implementation issues such as laplacian smoothing, numerical instability, etc. which they show you how to deal with in a quick and hacky manner. This lecture builds a full classifier from scratch in both the design and complete python implementation.

Machine Learning (Part 4 of 5): Semi-supervised Learning (Expectation Maximization)


This lecture builds on top of the Bayesian classifier that they developed last time. They build an expectation-maximization (EM) algorithm that locally maximizes the likelihood function. They go over an improved python implementation from the last lecture where they reduce the size of training set dramatically and measure the convergence of EM. This shows that using the testing data for improving the model greatly helps in improving the accuracy of the algorithm.

Machine Learning (Part 5 of 5): Unsupervised Learning (Gaussian Mixture Model)


This lecture builds a new python code from scratch. They use real valued numbers sampled from two different Gaussians with different priors. They build an expectation maximization algorithm very similar to the previous lecture. They also show how the algorithm converges to the two clusters without any labelled training data.