How does machine learning work?

Machine learning (ML) allows computers to come up with algorithms on their own for problems where it would be too complex for humans to formally specify a specific sequence of steps. This is done by giving computers free variables/parameters that they can adjust to reflect any patterns/features found in the data. When we combine this with a lot of data, it results in the machine being able to recognize underlying features of the data extremely accurately.

The industrial revolution automated tasks requiring physical solutions. The digital revolution automated tasks requiring mental solutions. We are currently living in the AI revolution which automates the finding of solutions itself. This means that now, we only need to formulate the problem and not the solution. One of the ways of automating the finding of solutions to a problem is by using machine learning.

At the onset of the field of AI in the 1950s, experts focused on representing problems using formal languages (logic, or programming languages). This was known as Symbolic AI. However, Real-world problems are often too complex for humans to be able to describe them using formal logic. This led to the birth of machine learning in the 1990s. This alternative method focused on letting the machine discover an algorithm and any underlying features of the data on its own instead of using experts to find hard-coded rules. This was done by giving the model a certain number of variables that it could change freely (free parameters). This allowed the models to learn how to represent the world on their own by manipulating these parameters.

Machine learning models can be categorized by the type of task we want the model to achieve :

An example of this is in machine vision systems. Before machine learning, human experts used to find and hand code features of images that they thought would be important for a machine to consider when trying to recognize an image e.g. horizontal or vertical line detectors. However as time goes on, and the diversity of objects to be recognized increases, relying on experts becomes slow and expensive. So we just let the machine learn which features are important on its own, instead of asking experts to pre-identify them.