Neurons (nerve cells) are the cells that make up the nervous system. They send and receive electrical signals to communicate with other neurons, playing a key role in distributing and storing information. The human brain is composed of hundreds of billions of neurons, allowing us to remember, make decisions, and experience emotions. While neurons consist of various components, we will focus only on the elements relevant to building artificial neural networks.
Neuron
In the context of artificial neural networks, a neuron is composed of dendrites, an axon, and an axon terminal. Dendrites receive information from other neurons and send it to the axon. The axon is a chain-like structure made up of multiple segments. The signal received by the dendrites is processed as it travels through the axon, sometimes increasing, decreasing, or disappearing. The processed signal from the axon is then transmitted to other neurons through the axon terminal.
Neuron and Artificial Neural Network
Let’s explain neural networks using the binary classification model discussed earlier. The input received by the dendrites corresponds to the input in a linear model. The signal processed through the axon represents the model. The signal processing involves weights (W) and bias (b), and the signal that passes through the model is shaped by an activation function to produce a specific result. The axon terminal, which sends signals to other neurons, corresponds to the output.
Single Artificial Neural Network
Using the linear regression and binary classification concepts learned earlier, we can build a simple neural network. A neural network consists of an Input Layer, Hidden Layer, Output Layer, and Back Propagation. The Input Layer, Hidden Layer, and Output Layer are similar to the concepts already discussed. Back Propagation is a method that applies the theory of minimizing the mean squared error between actual and predicted values. The difference is that errors propagate from the Output Layer back toward the Input Layer. This may not have a significant effect in a single-layer neural network but has remarkable effects in deep neural networks (those with multiple hidden layers).
Multidimensional Classification Problem
Let’s tackle a more complex problem: multidimensional classification instead of binary classification. A simple model cannot classify multidimensional data. Solving this problem requires tackling equations with hundreds or thousands of dimensions, for which deep neural networks are used, also known as deep learning.
To create a binary classification model, we first applied a linear model to make a primary prediction and then used an activation function to convert the prediction into non-linear data. Deep neural networks use this principle by dividing hidden layers into multiple nodes (models) that allow data to be classified through a combination of various non-linear functions. While the specific model choice is up to the analyst, the algorithm of the deep neural network automatically identifies weights and biases. With hundreds or thousands of hidden layers, astonishing results can be achieved.
Deep Neural Network
In deep neural networks, the input X may contain multiple values. For example, in a model that learns from images, an 800x600 image could be used as an input matrix of the same dimensions. In some deep learning algorithms, only a portion of the image is used as input.
In machine learning, inputs (X) and outputs (Y) can be processed in matrix form to handle various data types. Any data type, such as numbers, images, or audio, can be converted into matrices, and training is carried out through matrix operations.
The Hidden Layer consists of layers and nodes. Layers increase to the right, while nodes (circles) increase downward. The number of layers and nodes is determined by the analyst, largely depending on the performance of the computer used. As the number of layers and nodes increases, the computational power required grows, so it’s advisable to choose a level that does not hinder learning.
The most frequently used mathematical concepts in AI are basic arithmetic, logarithms, sequences, and matrix operations. Of course, those writing academic papers or conducting professional research will need to study mathematics more deeply, but for general AI learning, a high school level understanding of math is sufficient. High school math can be challenging, but you don’t need to master calculus in detail. If you encounter difficulties, a quick online search for key concepts will usually suffice.
