The machine learning process usually has 3 important steps.
1. The Creation of a model
2. The provision of initial input
3. Learning
To know how these steps work and how a machine really learns, we will need a real life problem that machine learning can solve.
Real Life Problem
You are a gym teacher who wants to know the hours your students need to train to lose the most weight.
1. Creation of a model
Machine learning normally starts with the creation of a model. A model is usually a prediction of identification that the machine uses to learn. A model must have some factors/parameters that the machine can use to calculate an output.
In our example, the gym teacher will tell the machine learning model to assume that training for 5 hours a day will lead to the loss of 10 pounds of weight in a month. In this example, the parameters the machine will use are the hours used for training and the pounds one will lose after training. The parameters can look something like this:
1. 0 hours = 0 pounds
2. 1 hours = 2 pounds
3. 2 hours = 4 pounds
4. 3 hours = 6 pounds
5. 4 hours = 8 pounds
6. 5 hours = 10 pounds
That information will automatically become - through machine conversion - a math equation since this is how machines express information. The equation mostly helps machines to form a trend that makes it easier for them to learn.
2. Providing input
Now the machine has model; the second thing it will need is an input. An input is real life information. The machine needs this information in order to see and learn a particular task works in real life.
In our example, the gym teacher will now need to feed the machine with information about how his students have been performing. For instance, he can key in something like this:
1. John 2 1/2 hours lost 3.5 pounds
2. John 4 hours lost 8.4 pounds
3. John 3 hours lost 5.3 pounds
4. John 1 3/4 hours lost 2.2 pounds
3. Learning
As you can notice, the gym teacher's input does not match the model he had given the machine. Some have fallen below the trend line while others have gone up the trend line. What happens when this occurs? Machine learning starts.
The machine uses the real life inputs also called training data to train itself on how to come up with a better model. The machine looks at the real life inputs to see how far off they are from the model. It then adjusts them using mathematical calculation to come up with a more accurate model. The model can change to this list.
1. 0 hours = 0 pounds
2. 1 hours = 0.8 pounds
3. 2 hours = 2.8 pounds
4. 3 hours = 4.8 pounds
5. 4 hours = 6.8 pounds
6. 5 hours = 8.8 pounds
With that said, for a machine to learn and be accurate, it needs more than one real life example. In our case, the gym teacher must feed the machine with another set of information and let it use it to adjust the model further. This cycle needs to go on until the teacher is out of inputs he had previously recorded. This way, the machine will be able to refine the model to a point where it will enable it to predict with ease the pounds one will lose when provided with the number of hours one works out in the gym.
With that stated, the above method is only one of the ways that a machine can use to learn. A machine can use three ways to learn. One of these ways is what we call supervised learning, which is what you have just learnt. The other two include unsupervised learning and reinforcement learning.
Unsupervised Machine Learning
In unsupervised learning, a machine normally learns - taught - with uncategorized and unlabeled data. Here, no teacher gives the machine input and output as we saw in the case of supervised learning. Here the machine just uses the data provided to mine rules, learn hidden patterns, and summarize data points that will help it determine an outcome and meaningful insight.
Unlike supervised learning, unsupervised learning is unpredictable. With that said, unsupervised learning has the ability to perform more complex tasks than supervised learning. One of the things that unsupervised learning can do is to identify pictures.
Here is how it can do that:
The principle is usually similar to supervised learning with the main difference being that the machine here does more work.
You first need to feed your machine with images, sa images of cats. When you do that, the machine will take over and then start building a model of likely images that can help it identify what a cat is in shape, colors, and images.
After that, the machine will now start learning and adjusting the images to come up with a clear vision and understanding of what a cat looks like. This process is usually difficult because it deals with object identification, which has parameters within parameters that the machine must use to translate the images into patterns. the patterns are what the machine uses to match objects.
Reinforcement Machine Learning
This method of machine learning uses observation that the machine gathers from its interaction with different things in the environment to figure out how to take actions that minimizes risk and maximizees its reward.
We call the reinforcement learning algorithms agents. These agents are what interact with the environment and then learn from those experiences with an aim of exploring the full possibilities of a state. In short, the agents are the ones that determine the ideal behavior that a machine can have within a specific setting.
In this process, the software agent is usually encouraged to learn the right behavior or action to take in a given situation through a reward system. A simple feedback normally works perfectly.
In reinforcement learning, the software agent has access to problems that it has to tackle by choosing the best action based on its current state. When the problem repeats itself, it gets a wiser software agent that has learnt how to deal with it and so the problem solving process becomes easier and faster for the agent. A problem that repeats itself is what we call a "Markov decision process," which is what the software agent relies on to improve its performance.
In short, reinforcement learning makes machines learn through experience. In this machine-learning model, the machine has to go through numerous problems, learn from them, and then do them perfectly the next time the problems come around.
A good example of AI inventions created by reinforcement learning is the game of chess. How did this happen? The designer programmed the machine with chess rules and then made the machine play hundreds if not thousands of games. The games gave it an opportunity to learn which moves it should make when on a certain setting.
Its learning process got a boost from the rewards the machine received after wining and the punishments when it lost. The machine stored those lessons in its teaching set and little by little, the program got smarter and better at playing chess. In short, the machine was "not coded" to output anything; instead, it "was made" to learn through doing a lot of practice.
These three methods are the methods your machine can use to learn how to mimic human behaviors. With that stated, these processes are not always perfect. Yes, they do their jobs, but they also have some limitations, like time constraints and algorithm limitations.
About the Author
Learning to teach machines to learn!
Are you intrigued by the fact that artificial intelligence poses an existential threat to human beings and has been around for at least 60 years? If yes, then here is the best introductory review of Artificial Intelligence and its effects on human behavior and the market. The book is thoroughly examined, neatly composed, significantly intriguing, and insightful. Help yourself understand the concepts of AI and get insights regarding:
• A brief history of artificial intelligence
• The state of art of machine learning
• Artificial neural networks applied to machine learning
• How to build an AI-ready culture
• Effects of AI on our daily lives
Reader April Williams says, "I brought this book for my AI class, surprisingly it's very interesting and an easy read. If you're interested in AI I recommend this book."
