AI Ladder: A Framework for Deploying AI in your Enterprise, Machine Learning. So machine learning algorithms, in general, are broken up into these four categories. First, we have supervised learning. Then we have unsupervised learning, deep learning, and then finally, reinforcement learning. Let's go into each one of these classes of machine learning algorithms. Supervised learning algorithms are the most widely used algorithms for prediction. A supervised learning algorithm learns from a labeled, pre-prepared, cleaned up training dataset. The algorithm is fed this labeled training data. And then that labeled training data is used by the algorithm to adjust its weights internally so that it can then make predictions when it's given new data that it's never seen before. So for example, you can use supervised learning algorithms for predicting a prospect's ability to repay a loan, predicting fraudulent insurance claims, or predicting machinery points of failure. Each one of these use cases can be accomplished by using predefined data that has been gathered historically, selecting it, and labeling it for all of the different outcomes you want the machine to recognize, and then feeding that predefined labeled data to the algorithm for its training. Unsupervised learning algorithms are most frequently used for anomaly detection and grouping data with natural affinities. So what you do is you take a bunch of data that is basically, not organized. You turn loose on to this data a trained algorithm whose job it is is to go through this unorganized data and group it and cluster it into the natural categories that it encounters. An unsupervised learning algorithm doesn't have an outcome variable. In other words, it doesn't make a prediction like supervised learning algorithms. Instead, it's used for segmenting data into groups with a natural relationship or a natural affinity for each other. So for example, you might want to use unsupervised learning algorithms for customer segmentation, to break up your customers into different categories, for detecting insurance fraud, for looking for those outliers that might indicate the presence of fraud. And they're most commonly used for image classification, for taking a picture and then classifying it into a particular category or bucket of categories based on the training and on the contents of the particular picture. Deep learning algorithms, you've probably heard about these in the news. They have been behind some of the big breakthroughs in machine learning lately. Deep learning algorithms are very efficient with working with huge amounts of data and unstructured data, like entire videos or movies or images. Deep learning algorithms are based on a family of algorithms called neural networks. A neural network is essentially a machine simulation of how neurons in the brain connect and communicate with each other. They're called deep because the algorithms consist of many layers of these artificial neurons. One of the big problems with deep learning algorithms is while they are very efficient with working with vast amounts of data, they're black boxes. In other words, once a deep learning algorithm does whatever it is you trained it to do, there's really no way that you can go in and figure out why the algorithm made the decision that it made. Because that decision is based on the architecture and configuration of millions and millions of artificial neurons that have all adjusted their weights to make some particular prediction. So deep learning algorithms are used today, for example, for making predictions in gene ontology and gene function relationships, the very complex nature of genetics. For making health predictions based on data collected from wearable devices. And action recognition in videos, recognizing certain acts in videos that should be called out and made available for others. Reinforcement learning algorithms, they are very commonly used where very large amounts of labeled data with the correct input output pairs are not explicitly presented. So typically, you've got an agent that is learning its way through some world by executing some action, and then learning whether or not that action was good or bad by getting a reinforcement or a reward or by getting a punishment that would prevent that agent from executing that action again. So for example, training a robot to learn specific policies. While the robot is engaging in different actions, sometimes random actions, the robot will receive either a reinforcement signal or a punishment signal to indicate that it has performed the correct action or not performed the correct action. Bidding in advertising can be used also in a reinforcement learning paradigm. And reinforcement learning paradigms are used a lot in games. For example, with non-playing characters, those non-playing characters may change their behavior based on the behavior of the playing character. And the behavior will alter depending on how that non-playing character is programmed to be rewarded for doing certain behaviors with the playing character or not rewarded for doing certain behaviors with that particular character.
