In the given magazine article, I would like to provide a simpler and plainer explanation of artificial neural networks to the general public by describing their current state of development. The target audience is comprised of individuals who are not educated and trained in the specified field, but willing to understand the overall structure and use of the technology. Neural networks represent a new and promising computational technology that provides new approaches to the study of dynamic problems in the financial field. Initially, neural networks opened up new possibilities in the area of pattern recognition. Later, statistical and artificial intelligence-based decision support and problem-solving aids in finance were added to this.
The ability to model nonlinear processes, work with noisy data, and adaptability makes it possible to use neural networks to solve a wide class of financial problems. In the past few years, an array of software systems has been built based on neural networks for use in a wide range of areas (Abiodun et al. 6). This includes operations in the commodity market, assessing the likelihood of bank bankruptcy, assessing creditworthiness, monitoring investments, and placing loans. Neural network applications cover a wide variety of areas of interest. It involves pattern recognition, noisy data processing, pattern augmentation, associative search, classification, optimization, forecasting, diagnostics, signal processing, abstraction, and process control. In addition, data segmentation, information compression, complex displays, complex process modeling, machine vision, and speech recognition can be distinguished.
Any neural network consists of an output layer and an input layer as well as neurons as units of the system. Independent and dependent variables are fed accordingly. The input data is transformed by the neurons of the network and compared with the output. If the deviation is greater than a given one, then the weights of the connections of neurons among themselves and the threshold values of neurons are changed in a special way. The process of calculating the output value and comparing it with the standard retakes place. If the deviations are less than the specified error, then the learning process stops. In addition to the output and input layers, there are so-called hidden layers in a multi-layer network. They are neurons that do not possess direct inputs of the original data but are connected only with the input of the output layer and the outputs of the input layer. Thus, hidden layers further transform information and add nonlinearity to the model.
If a single-layer neural network copes well with classification tasks since the output layer of neurons compares the values obtained from the previous layer with the threshold and outputs a value of either zero or one. Since it is not able to solve the most practical problems, a multilayer perceptron with sigmoid decision functions is able to approximate any functional dependence. However, in this case, neither the required number of layers nor the required number of hidden neurons, and the time necessary to induce complete learning is unknown. These problems are still faced by researchers and developers of neural networks.
The main difference and advantage of neural networks over classical means of forecasting and classification is their ability to learn. At the training stage, synoptic coefficients are calculated in the process of solving problems by the neural network. In this case, the required answer is determined not by the rules, but with the help of examples grouped into training sets. Thus, at the training stage, the neural network itself plays the role of an expert in the process of preparing data for building an expert system. The rules are assumed to be in the training data structure (Abiodun et al. 32). To train a neural network, training data is required, and it must meet the properties of representativeness and randomness or consistency. It is important to note the fact that everything depends on the class of the problem being solved. Such data represent a series of examples, indicating for each of them the value of the output parameter that it would be desirable to obtain. The actions that occur in this case can be called controlled learning, that is, the “teacher” feeds a vector of initial data to the network input and reports the desired value of the computation result to the output node.
Supervised learning of a neural network might be considered as an approach to an optimization problem. Its primary objective is to reduce the error aspect on a provided set of elements by choosing weight values. Since the error depends on the weights nonlinearly, it is impossible to obtain a solution in analytical form. More than a hundred different training algorithms have already been developed, differing from each other in optimization strategy and error criterion. The search for the global minimum is carried out through an iterative process of the so-called learning algorithm.
In sum, from the theorem on mapping almost any function using a multilayer neural network, it follows that a trained neural network is, in principle, capable of adapting itself to any. This is done in order to minimize the total squared error and to prevent this from happening, the following method of checking the network is used when training neural networks.
Reference
Abiodun, Oludare I., et al. “State-Of-The-Art in Artificial Neural Network Applications: A Survey.Hellionon, vol. 4, no. 11, 2018, pp. 1-41.