Nowadays there is a lot of talk about artificial intelligence, machine learning and deep neural networks, but few people understand what the difference is between different approaches to AI.
Artificial intelligence is actually at once several areas of computer science that solve problems inherent in the human mind: recognizing ¬ Speech, classification of objects, as well as various games like chess and go.
Machine learning is part of the topic of artificial intelligence, where it is not the direct programming of tasks that is studied, but the programming through the training -science in the process of solving one-type problems.
The concept of machine-building includes various algorithms, such as a random forest, decision trees, a naive Bayesian classifier , gradient busting and others. Neural networks, including deep ones, are also one of the algorithms for machine learning with artificial intelligence development company – Unicsoft.
Classic coding vs machine learning
How is the programming in the classical sense? Let’s admit that a person has a computer that works according to a certain algorithm. A person enters data into it, sets a program, and the algorithm produces results. In this case, everything is perfectly understandable. A person can get an accuracy of up to 100 percent, especially if the request is mathematical operations.
Scheme of classical programming
But in the case of machine learning, the program and outputs will change places. That is, a person gives the algorithm data and indicates the correct decisions, and then the computer thinks how to make it so that from these data the desired results were obtained. In the process of such work, a program is born.
The objectives of artificial intelligence
There are four main tasks of artificial intelligence:
Machine learning consists of two processes. The first one is training, when a person takes data, trains a model and as a result gets a classifier.
Machine Learning Model Training
The second process is already the use of ML, when the trained classifier is introduced into the system, and then new data is supplied to the input of the system. that the classifier has not seen. As a result, we receive predictions from the classifier.
Let’s consider an example of how a person uses his e-mail. It is possible to distinguish four patterns of human behavior, the analysis of which will help to determine its actions.
- At what time of the day a person uses mail: morning, afternoon, evening.
- How many devices it uses: a telephone, a computer, or at once several devices at the same time.
- In what locations is a person when he uses mail.
- In what order does a person check letters: top down or bottom up. We can determine this by how he replies or removes mailing lists and other rubbish from the box.
The totality of answers to these questions creates a portrait of a person (in the figure below, the behavior of such a person is highlighted in red). For machine learning, these actions will be predictable, without any bursts.
Now let’s imagine that the hacker hacked the mail, having somehow learned the password from it, and entered as a user. His behavior will clearly differ from the behavior of a person who used an electronic box before him. On the graph, the behavior of a hacker is shown with characteristic bursts. The task is to distinguish a cracker from a normal user by his behavior. For example, given: sites that a person visits, and the time the person is on them. It is necessary to determine the cracker after visiting the sites. The start page provides an overview of the problem and marked data. And also there is a tab where you can find how other users solved this problem.
What includes machine learning
Machine learning includes three components: data, signs and algorithms. Let’s take a look at them separately.
In open access, there are many datasets on which you can train the algorithms. But these sets have drawbacks. For example, sets may be incomplete, poorly labeled, and inaccurate. If you want to implement a solution based on ML technologies, you will need to collect a set of data for a specific task and the ready set is unlikely to match well. People are ready to lay out algorithms, tell what and how they use, but few people want to share their data sets.
Moreover, you will have to look for a compromise between these two parameters. For example, if the solution is to be used in real time, it is necessary that the model be able to read quickly.
Article by Born Realist