Artificial neural networks

I wonder if robots will rule the world and people, which is the subject of a hollywood movie, which attracts the attention of many of us in my column today. We will examine the artificial neural networks, which are the product of the blend of mathematics, medicine and informatics, which bring pregnancy to the question ”. It will be a long article, but besides the scope of the subject, today's article is related to its foundation and logic; will remain as a grain of sand within the scope of the subject.

Every baby is born with about 100 billion nerve cells responsible for brain functions such as thinking, seeing and feeling. In the later stages of life, no new ones are produced over these 100 billion innate nerve cells. Nerve cells are connected to each other by tiny gaps called synapses and form clusters that perform various functions of the brain. During the first eight months after birth, bond formation between nerve cells is surprisingly fast. After eight months, approximately 10,000 trillion synapses have been formed. Some of these synapses are preserved depending on the needs and stimuli in daily life, while the unused ones disappear over time and when the child reaches the age of 10, the number of synapses decreases to the level of adults (5,000 trillion). These connections formed in the first months of life are strengthened by external stimuli and made permanent. Those who give these stimuli are mostly the parents and the people who help in the care of the baby.

ANNs, which are simulated to the way the biological nervous system works, are networks formed by connecting neurons to each other in various ways. These networks have the ability to learn, store and reveal the relationship between data. In other words, ANNs realize the learning ability that a person normally has by living or experimenting.

Learning in human beings occurs by adjusting the synaptic connections between neurons. Neurons receive input signals through their numerous dendrites. An input by dendrites can be activating (trigger) or prohibitive. The inputs are collected and placed in the neuron body. When this input exceeds a certain threshold, the cell transmits an effect to other cells via its axon. This simple explanation creates a model for the artificial neuron. A neuron/processing unit is a processing element that collects incoming inputs and produces an output only when the sum of the inputs exceeds the internal threshold. There are five commonly used internal threshold functions. These are the liner, ramp, step, sigmoid and tanh(x) functions. As a threshold unit, it takes signals from neuron synapses and sums them all up. If the collected signal strength is strong enough to exceed the specified threshold, a signal is sent along the axon to stimulate other neurons and dendrites, and these signals recruit the cell body of the other neuron. The total signal is then compared with the inner threshold of the neuron and if it exceeds the threshold, it emits a signal to its own axon (Haykin, 1999). ANNs are mathematical systems created by connecting neurons working in this way and turning them into a network.

As a matter of fact, we can explain this shape with an example from our daily life, considering our biological structure: A baby wants to eat by holding a spoon. Until now, he knows that he has to eat when he is hungry, and he knows that he should eat the food from his mother, who is a reliable source, by recognizing his mother, that is, by seeing his face and hearing his voice, and he knows that he should swallow it. First time eating with a spoon! The desire to eat with a spoon comes from the desire to see and imitate his mother and the individuals around him. He's seen it before, experienced it, and taken it for granted. Well, a question for you: Is it possible for this baby to eat without spilling when using a spoon for the first time?

I'm sure everyone's answer was no. This answer was actually thanks to what you had learned using your nervous system before, the information that passed through your nervous system. So why should we ask?

One of the most important parts of ANN is training algorithms. The training algorithm specifies which type of learning rule to apply to the ANN according to the nature of the problem at hand. There are two types of learning rules; the instructional learning rule and the unsupervised learning rule. In the instructor-learning rule, the training algorithm is repeatedly adjusted using the input/output data until a convergence is achieved between neurons. In the unsupervised learning rule, ANN uses only input data and the relationships between them are expected to be learned by ANN by itself. In general, the most used training algorithm in ANN is instructor-led learning. In this type of learning, the ANN is trained before it is put into use.

During the training, both the inputs and the outputs to be created against those inputs are given to the ANN. In this way, the relationships between the inputs and outputs of the event are learned. The training phase usually takes quite a long time. People enter into a learning process by living and experimenting since their birth. In this process, the brain is constantly developing. As we live and experience, synaptic connections are adjusted and even new connections are formed. In this way, learning takes place.

This also applies to ANN. In order to create a learning rule in ANN, input/output data may need to be introduced to the network many times. In today's education and training systems, research on obtaining information by making use of the effect of environmental stimuli on the sense organs is at the forefront. In the ANN system, information comes with stimuli in the external environment. In order for information to be stored in memory as in the brain, meaningful coding is required. In addition, the coded information must be brought back in order for learning to take place. Learning must occur before information is stored in the brain. For learning to take place, the number of connections between neurons must be high. The greater the number of networks formed by neurons, the greater the learning. Learning in the educational process is a relatively permanent change in behavior that occurs as a result of the interaction of the individual with the environment at a certain level. Learning occurs as a result of the active use of neurons. Considering in terms of education, the high number of neurons is important in terms of the learned knowledge and the permanence of the knowledge.

In short, in order for the baby to learn to eat with a spoon, he needs to try many times and train the neural network until he reaches the real result.