Artificial intelligence has undoubtedly led to significant revolutionary changes in today's software and hardware technologies. In fact, although this term has become popular in recent years, artificial intelligence studies actually date back to the 1950s. One of the most important reasons why the term artificial intelligence has gained such popularity in recent years is that the application areas and performance have increased greatly and rapidly with the developing new algorithmic techniques. Due to these developments, artificial intelligence technology has led to the replacement of old solutions with new generation solutions by overthrowing the solutions in many fields, like dominoes. The increase in the use of artificial intelligence technology has led to the introduction of new terms into the literature, and this has caused an important conceptual confusion. Artificial intelligence, machine learning, artificial neural networks and deep learning can be given to the terms that are frequently encountered in this confusion today. So what is the real meaning of these terms? Let's try to shed some light on this issue together.
Artificial Intelligence vs. Machine Learning
Artificial intelligence is a field of computer science that aims to develop 'intelligent' machines that can act (think and make decisions) like humans. Its aim is to develop autonomous systems that can make decisions by imitating people's intelligence and successfully perform a specific task like humans. This includes the ability to imitate abilities such as perception, learning, and causation, which are constantly used by humans, by artificial systems.
Machine learning is about machines; It can be thought of as learning by making use of the past data provided without the need for human input, making predictions about the future with what they have learned, and adapting according to changing data. In this context, machine learning is positioned as a sub-branch of artificial intelligence. Machine learning techniques are divided into three main categories according to the methods used in the learning process:
- supervised learning
- unsupervised learning
- reinforcement learning
The difference between artificial intelligence and machine learning can be made more concrete with an example. For example, let's say we are developing a bot for the game of checkers. Let's say we have prepared an algorithm for the first version of this bot. Let's pre-program the moves of the bot according to the game rules and scenarios that will improve winning. For example, we can use symbolic logic algorithms and decision-making mechanisms for this. In this way, we can evaluate systems that can make decisions according to a pre-programmed logic and therefore imitate human intelligence under the umbrella of artificial intelligence. However, the first version of the bot does not include any machine learning steps. Let's plan a second version of our bot like this. Let's take a data record of the moves made and the results of the moves in the previously played checkers games. Let the algorithm decide which move would be most beneficial according to the scenario it is in, based on the data provided during the learning phase. In this case, the developed bot can be qualified as an example of machine learning. Because in this scenario, the bot learns how to play the game without the need for any human input, except for the data provided during the learning phase. Of course, since machine learning is a sub-branch of artificial intelligence, our bot in the second scenario is also an example of artificial intelligence.
Types of Machine Learning
Supervised learning is one of the techniques frequently used in machine learning. In this learning style, results labeled according to ready inputs are used as training data. Thus, the learning algorithm performs the learning process according to known input and result relationships. It then tries to predict the most appropriate result for a given test data. The classification problem can be given as one of the best examples of this situation. For example, let's say we want to distinguish animals from each other in a dataset with pictures of chickens, cats, and spiders. In this context, the supervised learning algorithm first tries to reveal some features that will distinguish chickens, cats and spiders from each other by using pictures that have been tagged by people before. Thus, between these three categories, it tries to determine the most defining features that will enable the elements to be separated from each other. Then, when a test picture is given, it tries to determine which category the test picture is closest to, by applying the same operations to the attributes in that picture. Thus, the supervised learning algorithm executes class prediction on the test picture.
Unsupervised learning, on the other hand, tries to understand the similar and distinctive features of the input items without any previously labeled outcome data. One of the best examples of this situation is the clustering task. If we reconsider our previous example, this time
It is intended to classify a set of images of chickens, cats, and spiders without the pre-labeled dataset described. It is not even known that the pictures given in this scenario came from these three animals. In this context, the number of classes to be obtained as a result is also uncertain.
Finally, reinforcement learning can be thought of as a perfect synthesis of supervised and unsupervised learning. In this context, in the reinforcement learning method, the algorithm makes an exploratory decision for the incoming input. It associates the right or wrong decision with a certain reward function. Thus, the algorithm decides what to do with the new test signals with the experience gained from the right and wrong decisions. In this respect, reinforcement learning can be thought of as a synthesis of guessing and random discovery based on existing knowledge. One of the best examples of this situation is the hot and cold game we played as children. The competitor, who has no prior knowledge of the location of the sought target, first proceeds in a direction for exploration. He tries to find the target based on the instant hot and cold feedback he receives from other competitors and his experiences in the competition.
Artificial neural networks
Artificial neural networks are a machine learning technique that is modeled from the biological learning structures of humans. Learning and decision-making mechanisms in humans are carried out by neural networks consisting of neurons. The aim here is primarily to model neurons, which are the building blocks of this learning system, with various mathematical functions. Then, with the help of networks formed from these artificial nerve cells, some tasks such as classification, clustering and learning that humans can easily do can be performed by machines.
In this respect, artificial neural networks appear as a sub-branch of machine learning. A significant majority of artificial neural network techniques used in the literature can be evaluated in the supervised learning category. That is, neural networks are trained using manually marked results against inputs. Then, the artificial neural network structure that emerged as a result of this training is tested on new test inputs. However, there are artificial neural networks that can operate unsupervised. One of the best examples of these are self-organizing networks, also known as Kohonen networks. In these networks, the aim is not to update the network function in a way that minimizes a certain cost function, but to cluster the given high-dimensional inputs in a way that can express them in a lower dimension.
Deep learning is a machine learning technique that aims to enable people to perform critical activities such as information processing and decision making. Deep learning algorithms are basically based on artificial neural networks. However, unlike artificial neural networks, the deep learning algorithm can automatically create as many hidden layers as necessary. In fact, the term 'deep' comes from the depth of the number of hidden layers.
Deep learning algorithms can learn the relationship between input and result by themselves, as they can contain many hidden layers. That is, the attributes do not need to be specified beforehand, as in many classical machine learning techniques. Of course, this may sometimes cause algorithms to use different attributes that they have determined instead of the targeted attribute. Consider the example of the chicken, cat, and spider we gave earlier. Our aim here is to recognize and distinguish these three animals from each other based on certain attributes. In such cases, classically, certain features related to the faces of animals are determined and the machine learning algorithm is expected to distinguish animals from each other based on these features. However, when the same sample is processed with a deep learning technique, classification or clustering can be done without determining these features. However, in this case, the deep learning algorithm will determine the distinctive features itself. One of the important distinguishing features for this scenario is the number of legs of the animals used in the training phase. In other words, the deep learning algorithm can make this distinction over an attribute that we do not actually want or plan, without being aware of it.