Machine Learning Demystified

Machine Learning is all the rage these days in computing. But what exactly is it? It is a subset, or part of Artificial Intelligence. Machine learning is a way of giving some intelligence to computers. Just how much depends on how well the machine can learn, which is in large part based on what it is taught. But how can the machine be taught?

"The majority of practical machine learning uses supervised learning." ¹

First, we can consider the two major categories of Machine Learning: supervised and unsupervised. In supervised learning, inputs and desired outputs are known. But what are these inputs and outputs? One very basic example is image identification. Being able to see if an image is a cat, a dog, a baby, etc. The inputs of the training data are the pictures of these things. The outputs are the correct word labels of "cat", "dog", "baby" that the algorithm will associate with the pictures.

The algorithm here is the programmatic process by which the "machine" is learning. The machine here is usually a computer very much like any we are familiar with, and the algorithm is just some kind of software program that runs and processes the training data. This process is mathematical and statistical in nature, crunching numbers along the way to calculate how similar one thing is to another based on various measurements.

The training data is basically the set of examples by which the machine learns. In this case the training data is the set of pictures and the pre-determined correct labels associate with each one. As this training data is processed the algorithm becomes more effective at being able to correctly identify new pictures beyond the training data. Of course in our case it will be only trained to identify the objects it was supervised to identify, such as the cat, dog, or baby.

One way to imagine how the algorithm becomes better at identifying correct answers over time is that it plots the input-outputs, or questions-and-answers of the training examples on a graph. And then it tries to find different lines or curves that fit the points as best they can. As it processes more training data, and plots more points, and refines its graph to fit those points, that graph will become better at predicting what the answer should be to new questions. Such as, "does this new picture contain a dog?"

In unsupervised learning, the desired outputs are not known. So in the case of pictures, you might have many pictures of different kinds of animals but the animals in the pictures are not identified. So the machine is allowed to learn without "supervision" what the pictures contains. The way it does that is by finding subgroups in the group of pictures that share common characteristics. So it might quickly learn to identify a new type of animal, a giraffe, because the length of its neck is consistently longer than the other type of animals. Again, the way these comparison are done by the machine is numerically, subtracting the value of one thing from another to see how much it differs by. Doing this for various values, and drawing conclusion from the quantitative differences.

"Evaluated with respect to known knowledge, an uninformed (unsupervised) method will easily be outperformed by other supervised methods, while... supervised methods cannot be used due to the unavailability of training data." ²

Two common quantitative methods are classification and regression. Classification is used when the data is limited to a finite amount of values. Such as a set of categories of animals. Regression is used when the data is continuous, that is can take almost any value, such as predicting tomorrow's price of milk based on past prices. Although technically since money cannot be divided below a cent this is not continuous. So to be technically correct we could consider the tomorrow's temperature, based on historical data, since that can take any decimal value.

"How is neocortex relevant to machine learning?" ³

Combining supervised and unsupervised training methods can be very effective. Starting the machine off by teaching it some basic animals to identify and then having it search through new pictures to learn to identify new animals that it has never been taught about before. In this sense we can start to see how even an artificial machine like a computer could start developing a type of intelligence in ways very similar to humans. In fact, it is very possible that our own minds do something very similar to "machine learning."

1. https://en.wikipedia.org/wiki/Machine_learning

2. https://www.geeksforgeeks.org/regression-classification-supervised-machine-learning/

3. https://towardsdatascience.com/machine-learning-of-human-brain-739ab0419612