May 27, 2020 by Andrei Demit
Demystifying Machine Learning: Part I
A Three-Part Series on Machine Learning Techniques
Machine Learning (ML) and Artificial Intelligence (AI) are buzzwords heard everywhere today, especially in the news about technology. It seems as everything today has some AI and ML properties. Marketers have even begun to use these terms to make their goods or services appear more attractive, making ML and AI such ubiquitous acronyms that it seems you can find them in the most unlikely of products.
Some people debate whether it is more of a marketing tactic or if AI and ML are everywhere. To assume a position on that, we need first to understand what are AI and ML.
Artificial Intelligence is intelligence exhibited by machines in terms of learning and problem-solving capabilities. More broadly, AI is any manifestation of action that belongs to the conscious biological creature’s reign. Furthermore, the human brain loves to perceive AI as something magical. Therefore, once it understands the inner workings, the magic fades away, and the definition of real AI changes again by setting the capabilities a real AI should have even higher.
Machine Learning, instead, is a subfield of the Artificial Intelligence umbrella that aims to achieve AI through learning and pattern discovery in large datasets. Those patterns are essential to make predictions, recognize elements, behaviors, situations, and make choices that are considered appropriate (based on past data).
Further, Machine Learning divides into three main approaches: Supervised Learning, Unsupervised Learning, and Reinforcement Learning.
Let’s start by analyzing the first of the three categories into which ML divides– the Supervised Learning method.
As the name implies, this learning technique makes use of a supervisor(usually a human), which tells the machine what the outcome of a series of variables is. By providing enough examples, the machine learns to give a weight to each variable to predict the result. This weight is the amount of importance a specific variable has in influencing the outcome.
From a mathematical perspective, this learning implies solving an equation. We have f(x) = Y where (x) are the variables and Y the result. What the algorithm has to find is f(x), which represents the function that best maps the (x) to the Y.
As you may infer, Y is always a number; thus, all the variables (x) are also numbers. However, the value of Y has the potential to represent two kinds of results, one is the result of a regression function, and one is the result of a classification function.
A regression result aims to represent the value of something. “Something” takes many forms, such as product price, stock price, temperature degrees, and the number of customers. To determine whether the issue you’re trying to solve requires a regression approach, try and see if your solution can be the answer to questions, like:
- What temperature will be tomorrow?
- What is the value of this diamond?
- How many customers will buy this product?
- How many cakes should I prepare today?
If yes, then you have a problem that has the answer in a supervised learning – regression approach.
On the other hand, a classification result aims to assign a class to something. In this case, “something” can represent, through a value comprised between 0 and 1, the likelihood an image depicts a dog or a sheep. It can refer to whether a transaction is legitimate or not or whether a spot on the skin is melanoma. The comparisons pairs are limitless.
But the real power resides in the way this can be combined and used inside neural networks. By putting together many so-called artificial neurons, where each one of them recognizes a part of an image’s features, it is possible to create a machine that identifies and classifies objects.
As with the regression, the classification also has a series of questions it can answer:
- Is this an apple?
- Will the stock price rise, or will it fall?
- Is this email spam?
The majority of AI out there is using this class of learning, Supervised learning. That is because the majority of applications are dealing with predictions and classification of things.
Real-world use cases, like object detection, image classification, fraud detection, forecasting, diagnostics, all make use of supervised-trained ML models.
In this article, we have gone through the high-level concepts of AI, ML, and specifically Supervised learning. Machine Learning has two more learning techniques it can handle: Unsupervised Learning and Reinforcement Learning. We will explore each one of them in separate future articles, as well as provide a more detailed explanation about algorithms used in each type of learning.