Forecasting.ai

Forecasting AI blog

Blog

>

Machine Learning

>

Which machine learning system to choose?

Which machine learning system to choose?

By G. H.

|

April 14, 2022

|

Machine Learning

|

enfr

Recent developments in AI (artificial intelligence) technology have led to a number of breakthroughs and exponential growth in machines. The extent to which the entire world is now dependent on machines knows no bounds. In fact, AI solutions are now not only a key investment opportunity for large companies, but they are also a major contributor to solving countless everyday problems in our lives.


One of the most important subsets of AI is machine learning, often referred to simply as ML. It is only because of the invaluable work that researchers and scientists have invested in the foundations of ML that we are now able to extract maximum performance from highly skilled AI-based technologies.

In this article, we'll look at how, over the years, humans have made machines capable of intelligence, that is, capable of mimicking the human thought process and making decisions based on their experiences.

What is machine learning?


Before we talk about the different methodologies humans use to teach machines to behave like humans, let's review the basic definition of machine learning.

Machine learning is the method by which humans teach machines to learn from a set of past data and allow them to perform certain actions in the future based on the past learning. Machine learning is a combination of many elements, from computer algorithms to data analysis to mathematics and statistics. It is the technology on which the construction of artificially intelligent systems is largely based.

How are machines trained?


The process by which machines learn from previous data is called training.

The science of machine learning involves teaching a machine by using datasets containing useful or random facts and/or data of various sizes and feeding them into the machine. The essence of this activity is to help the machine observe the data, make meaningful connections between the various pieces of information provided, and train itself to make decisions about the incoming data by incorporating these preconceived relationships, also called rules.

Machine learning models often follow one or more of the following main training methods.

  • Supervised learning
  • Unsupervised learning
  • Reinforcement learning


For initial training, a data set is used whose expected input and/or output may or may not be explicitly defined. The training process uses training data. Once the machine is trained, it receives test data to see whether or not it has learned from the trained data set.

Let's review these training methods in a bit more detail and look at how they are used to make machines smarter.

Supervised learning


This type of machine learning algorithm uses a dataset that contains labeled data. This means that it tells the machine what each item is. So in theory, we can predefine the rules and all the machine has to do is study the existing matches and learn these rules.

Supervised learning algorithms can be divided into two subtypes: classification and regression.

Classification Classification: This method is used when the machine has to be trained on binary answers such as yes-no, right-wrong or true-false. The training data consists of items that have already been classified into different categories. For each category, the machine carefully examines each item and identifies the characteristics that are unique to all items in that category. This allows the machine to construct relationships between items and their corresponding categories. It uses these rules to identify and correctly classify items in the test data.

Regression A regression model is used when predictions are needed for numerical values such as house prices or temperature. The training data set contains several variables and outputs that may or may not depend on these variables. The machine studies the input variables and determines how each variable influences, if at all, the value of the output, leading to pattern recognition or rule development. For test data, the machine uses these rules to calculate the estimated or predicted value of the output.

Unsupervised learning


The main difference between supervised and unsupervised learning is that in the latter dataset, the items are not labeled. Let's take an example to better illustrate this point.

Suppose we want a machine to be able to classify items in a dataset containing images of different types of gardening tools, such as hoes, shovels, rakes and spades.

In the case of supervised learning, the training data contains images with their identifiers. For example, if you enter an image of a shovel, the machine will know that it is a shovel. The machine will then study all shovels and their common characteristics to learn how to identify a shovel in the future.

However, if you use the unsupervised learning model, you will input images of all kinds of gardening tools without labeling them. For example, if you input an image of a shovel, the machine will not know that it is a shovel. The machine will have to figure out for itself how each image can (or cannot) be related to the one before it, and then group similar images into a category. So the machine learns to create categories on its own without being explicitly told what those categories are. This type of training model works well for datasets whose structures or patterns are not necessarily obvious to the average person.

Reinforcement learning


The third major method is based on the concept of reinforcement, which some of you may already be familiar with if you've ever taken a Psychology 101 course. If you've ever tried to teach your dog cool stuff by motivating him with reward walls, you've used the reward system.

Unlike the first two methods, this model relies heavily on feedback. For each decision, the machine reports the correct outcome to see if it made a good or bad prediction. Through repeated trial and error, the machine becomes more accurate.


A concrete and simple example of reinforcement learning can be seen in the display of online ads. The machine can determine which ad is more successful and worth displaying based on the number of clicks. If the machine gets more clicks (higher rewards) from a particular target group for a particular ad, it will know that it was a good decision to display that ad to that group.

Final thoughts


While some seem determined to settle the human/machine debate once and for all, others feel that this type of comparison is futile. The fact is that man came first and the machine came second. As long as our passion for growth and need for perfection lives on, machine learning algorithms will continue to evolve and become more accurate, helping us achieve success and accuracy rates that seem impossible.