What are Artificial Neural Networks?

Artificial neural networks (YSA) have become one of the most popular machine learning algorithms of the last 15-20 years. Inspired by the working principles of the human brain, these networks are considered one of the most powerful machine learning algorithms of our time. If we answer the question of artificial neural networks, artificial neural networks are systems that gain the ability to learn by mathematically modeling the human brain and can create patterns of solutions to different problems autonomously. To learn more in depth about the structure of artificial neural networks, let's first take a brief look at how the human nervous system works and its structure.

Structure of Neuron

The main cells of the human nervous system are neurons. These neurons process the signals we receive through the senses and transmit them to other neurons. Through dendrites found in neurons, signals from the axon ends of other neurons are transported to the cell nucleus and these signals are processed in the cell nucleus and transmitted to the axon end of the neuron. Finally, signals are transmitted from the axon end to other neurons. The central nervous system consists of over 100 billion neurons.

Structure of Artificial Neuron

After briefly looking at the structure of neurons in the human brain and the logic of operation, we can examine the structure of artificial neurons that we create on computers. Each neuron consists of inputs, weights, threshold values, activation function and outputs. We can compare inputs to signals received by human senses, activation function to cell nucleus, and output to axon tip. Artificial neurons process incoming inputs just like human neuron cells and transfer them to the next neurons. This process is calculated according to the formula above. Because each entry is multiplied by weight values according to the formula, weight values indicate the importance of input values. Inputs with a low weight value contribute less to the output value obtained in the latest artificial neural network, while inputs with a high weight value contribute more to the output value. This value is then passed through the activation function and the value to be transferred to the next neurons is calculated. For the moment, these activation functions may remain a question mark for you, and in our future articles we will conduct more detailed examinations of the contents of these functions and why they are used.

Structure of Artificial Neural Networks

Just as neuron cells come together to form the human nervous system, artificial neurons form artificial neural networks through their connections. Artificial neural networks consist of one input layer, one or more hidden layers, and one output layer. Artificial neural networks are trained on the appropriate data sets for the desired problem. The goal of this training is to find the most appropriate values for the parameters in the YSA model. These parameters are the weight values and threshold values of neurons. Throughout the training, our parameters are constantly updated. We perform this update according to the loss function of our model. The loss function is a criterion for the success of our model on the data set. Our goal is to find parameter values that minimize this loss function, that is, bring it closer to 0. As an example of a loss function, we can give an mean squared error, which is one of the most commonly used functions. The average frame error is an indication of how close or far away our model's forecast value is for each instance in the data set. According to the formula, each data set sample is squared by the difference in our prediction. Frame retrieval is done both to get rid of negative values and to give more contribution to examples with a high margin of error. Then each margin of error is collected and divided by the number of samples, averaged. After the value of our loss function is calculated, we minimize our loss function with certain optimization techniques. Being able to choose the right loss function for the purpose of our model and using the appropriate optimization techniques has a great effect on the success share of our model. Therefore, in our following articles, we will examine various loss functions and optimization techniques in more detail.

Areas of Use in Our Daily Lives

After taking a look at the structure of artificial neural networks and how they work, we can take a look at their usage areas in daily life, where they appear. Natural Language Processing: Examples of this field include automatically generated subtitles on Youtube videos, and automatically classifying user reviews as positive or negative on e-commerce sites. Computer Vision: Examples include autonomous Tesla vehicles in the automotive industry. During driving, the recognition and classification of many objects is carried out. The location of many objects such as pedestrians, traffic lights, other cars is detected. In the health sector, certain diseases are detected according to radiological images. You can check out our article On the Victory of Artificial Intelligence Against Koranavirus. Art: Artificial neural network models can transfer these analyses to new images by identifying lighting, brush strokes, colors and finding patterns between them. As another example, artificial intelligence models can be trained on songs composed by groups and compose new and original songs as if they had composed them. For example, you can listen to Nirvana – Drowned in The Sun, composed by artificial intelligence 26 years after the death of Kurt Cobain. Video Games: Fifa, a football game, is an example. When managing our team, we can only control one of our players, but it's thanks to AI models that our other players make sensible decisions based on the course of the game. Banking: Banks use artificial neural network models to test fraudsters, conduct credit analysis and automate financial advisor systems.