Introduction to Machine Learning, Part 2
Welcome back, folks, to my series of blogs summarizing what I learned about Machine Learning from the For Dummies series (please see the full source listing at the bottom of this post). Please note that all the information presented in this post is from Hurwitz and Kirsch, except where otherwise cited.
Source: https://www.em360tech.com/tech-news/anything-you-can-do-ai-can-do-better-machine-learning-in-insurance/
The first concept I want to highlight is about where machine learning fits in with artificial intelligence. According to Hurwitz and Kirsch, machine learning is a subset of artificial intelligence; other subsets include reasoning (making "inferences based on data", Hurwitz and Kirsch, pages 13), natural language processing (training computers to understand text and speech) and planning (allowing a computer to independently develop a sequence of actions to meet an objective) [1].
Within machine learning itself, there are also different categories of learning that can be applied to a data set, depending on the characteristics of that particular data set. The 4 key types of learning are [1]:
Supervised learning: this learning can be done with a data set that is well understood and labeled. Supervised learning is used to find trends in data sets. For example, the data set could be a collection of thousands of images of animals that are labeled so that the algorithms can find similarities between different species. If the labels used to characterize the data are continuous, then regressions are required to analyze them. If the labels are discrete, then just a simple classification algorithm is required [1]. Regression algorithms are statistical methods for estimating the relationships between variables [2].
Unsupervised learning: this method works best with unlabeled data sets, such as data collected from social media sites. The mathematical algorithms that fall in this category are designed to look for patterns or clusters (i.e. groups of data with similar features) in an unlabeled data set. The algorithms work iteratively until they have classified the data satisfactorily. Often unsupervised learning will be used to label an unlabeled data set before it is passed on to a supervised learning process. It is used today in e-mail spam detection to identify spam e-mails [1].
Reinforcement learning: this method does not use a training data set, but instead learns through trial-and-error. Robots can use reinforcement learning to develop a series of actions that successfully achieve an objective. If the robot gets positive feedback on an action, the action is repeated; if the feedback is negative, the robot adjusts the action in an attempt to improve the outcome. Today, reinforcement learning is used to develop the AI in self-driving cars [1].
Neural networks and deep learning: deep learning is the fourth category of machine learning, and it leverages a new technology called neural networks to operate. Neural networks are designed in layers to mimic the way the human brain works. Neural networks are intended to be able to deal with difficult real-life problems that cannot be well defined, like image recognition, speech and computer vision. Neural networks are made up of nodes in a sandwich configuration: there is an input layer, a series of one or more hidden layers, and an output layer. Data is passed from one hidden layer to another based on the weight assigned to each node. Deep learning refers to the process that occurs when multiple hidden layers are used to solve a problem. Deep learning learns from unstructured data [1].
A note - supervised learning depends on the data sets that are used to "train" the mathematical model. Once the model has been trained on the training data and is then tested on a new data set, it may not be able to find the same patterns that it could find in the training data sets. If the model really only works with the training data, this is called "overfitting," because the mathematical model has been fit too tightly to the training data, and now it cannot identify patterns in a broader data set [1].
In my third and final post in this particular series, I will review some different types of machine learning algorithms.
[1] Hurwitz, J and Kirsch, D. 2018. Machine Learning for Dummies: IBM Limited Edition. New Jersey (NJ): John Wiley & Sons, Inc.
[2] https://en.wikipedia.org/wiki/Regression_analysis