How do Artificial Neural Networks Work

An artificial neural network is one of the most promising new machine learning technologies that has evolved in recent years as AI has gained acceptance in the academic community. The human nervous system serves as a model for the development of artificial neural networks. When an organism is triggered from the outside, messages are conveyed by neurons, and the organism's awareness is engaged to generate relevant feedback. This occurs because the biological neuronal cells are linked by synaptic connections and constitute complex neural network systems. An artificial neural network mimics this quality by linking together many virtual neurons, exchanging data through the right activation functions, and processing data in a manner similar to that of a human brain. (Hinton G.E., N. Srivastava, A. Krizhevsky, and E. et al. There are three distinct ways of thinking that may occur in a human brain: abstract, picture, and inspiration. Image thinking, in particular, is the exact direction of AI development, as it allows machines to learn knowledge and produce subjective cognition that can replace human work by inspiring creative solutions to problems. To aid or perhaps replace humans in many jobs that require subjective identification or prediction, artificial neural networks have shown to be an excellent tool. Deep neural networks and convolutional neural networks are only two examples of the state-of-the-art models and algorithms that have emerged as a result of the rapid advancement of artificial neural networks in the early 21st century. Simultaneously, there is a clear split between theoretical and practical orientations in terms of the artificial neural network assessment indices. An algorithm's theoretical impact may be measured by how well it performs in pattern recognition or prediction.2015) (C. Szegedy, W. Liu, Y. Jia, et al.) Furthermore, AI is in its formative years, and building a comprehensive theoretical framework is essential for the advancement of artificial neural network theory. As the complexity of the neural network model rises, so does the volume of data it must handle; as a result, researchers are also interested in optimizing the operating overhead and learning speed of neural networks for practical use. Before analyzing the major research outcomes of artificial neural networks, this study examines the fundamentals of neural networks from both a theoretical and an applied perspective.

An artificial neural network is a model with several interconnected layers, each representing a different kind of neuron. The output value of a neuron in a layer that is not the input layer is calculated using the activation function of the neuron with which it is linked in the lower layer. The vector of input parameters is read by the neurons in the input layer, which then sends its output to the neurons in the following layer. Neurons employed for processing input and extracting features have extensive interconnections that are hidden from view in "hidden layers." The categorization outcomes or estimated values might be produced by a single or several neurons in the output layer. A connection between input data and output data is established at a single neuron node. The activation function f, the weight parameter W, and the output value hw,b(x) are all defined by the following equation: hw,b(x) = f (WT x) = f (∑ 3i=1 Wi xi + b) where x1, x2, and x3 are the input values and +1 is used to introduce bias value b to set threshold for the activation function. Nonlinear functions, such as the Sigmoid function, the Rectified linear rectifier unit function, etc., are used to activate the input data when confronted with complicated nonlinear classification jobs and the like. Layers of input data, hidden processing layers,

of the input layer, the activation function is used to determine the output value of any neuron in the network by multiplying its input data by the output data from the connected neuron in the low layer and its weight. Training a neural network mostly entails fine-tuning the network's weight parameters in order to achieve optimum performance. The practical application impact of an artificial neural network requires training and learning using massive data. Back Propagation, also known as the BP algorithm (Glorot Xavier, A. Bordes, and Y. Bengio. 2011), is now the most used training technique for neural networks.The BP algorithm is implemented by first feeding the neural network input from the training set, then using forward propagation to calculate the network's output values, and finally using a mean-square cost function to minimize the difference between the network's output and the expected output value of the training set. Finally, the network's weight parameters are adjusted backwards, layer by layer, using the Gradient Descent Algorithm. After many rounds of training, a neural network is termed convergent if the error between the network output value and the predicted value stays within a limited, stable value range.

WANG Xin-gang (2017) with the advancement of AI, there has been a heightened interest in artificial neural network algorithm models that mimic the activity of human brain neurons. Artificial neural networks' feature extraction and classification abilities have found widespread application in areas such as image recognition, speech recognition, natural language processing, trend prediction, and others where human presents the analytical thoughts and views by analysing and comparing numerous neural network technology models and algorithms based on the development history and research state of artificial neural networks. Enzo Grossi, Massimo Buscema (2008) Coupling computer science with other theoretical foundations like nonlinear dynamics and chaos theory paves the way for the development of 'intelligent' agents like ANNs that can dynamically adapt to issues with a high degree of complexity. The study of complexity is made possible by the fact that ANNs can mimic the dynamic interplay of several components concurrently, and they can also make conclusions on an individual basis as opposed to only observing patterns. When compared to traditional statistical methods, these tools may have some distinct benefits. The goal of this paper is to introduce gastroenterologists to ANN-related ideas and paradigms. When applied correctly, the ANN family allows for the greatest possible insight into complex, dynamic, and multidimensional events that are difficult to forecast using the more conventional "cause and effect" framework. Nitin Malik (2005) The purpose of an ANN, which is a functional imitation of a simplified model of biological neurons, is to build useful 'computers' for solving real-world problems and to reproduce intelligent data evaluation techniques like pattern recognition, classification, and generalization using simple, distributed, and robust processing units called artificial neurons. Non-linear static-dynamic systems are implemented in ANNs in a fine-grained parallel fashion. Because of its massively parallel and dispersed structure, neurons have a great processing capacity, which contributes to ANN's intelligence and allows it to handle challenging tasks. The ability to implement ANN algorithms and architectures with VLSI technology for real time applications is a major factor in the recent resurgence of interest in ANN. Theoretical and practical results in a wide range of fields have fueled a rapid expansion in the number of ANN applications in recent years. This article provides an overview of recent studies and the rapid expansion of several ANN-related applications. This article provides a high-level summary of ANN concepts, models, and uses. We explore future trends and identify potential application areas.

The process through which the published works were evaluated and analyzed is described below. Depending on the desired level of detail, several types of documents were gathered and analyzed. The reviews are categorized according to the kind of modifications made, the outcomes, the fields of published. Journals, conference proceedings, and technical reports were all used in the publishing process. When looking for articles on a certain subject, the term "artificial neural network using conjugate gradient" is often used as a keyword. To further illustrate the breadth of ANN's practical use, a second set of papers was selected. Therefore, the purpose of this work is to zero in on the most promising unexplored element of the development of CG formulation in ANN.

Since its discovery in the 1950s, ANN research has seen significant improvement and evolution. There are several methods in ANN that may be used to boost learning performance, so you aren't limited to the CG formulation. Many scientists, nevertheless, use different bio-inspired algorithm methods, such as the genetic algorithm, particle swarm optimization, bee colony optimization, and so on. The following table details the issues faced by ANN researchers, along with the methods they used to address them. Each study's findings are shown as well, demonstrating how the implemented changes enhance the experimental testing data's performance. The table also includes examples of the many different CG that have been used to solve ANN issues. Both feed-forward and back-propagation algorithms are used in this process. The results of using CG demonstrate an increase in precision, reliability, error reduction, and convergence speed.

There has been significant development in several study fields thanks to ANN. ANN is well-suited for use in the fields of medicine and forecasting. More and more fields, from medicine and mathematics to commerce and engineering, are finding usefulness in this expanding field. Figure 1 shows that the fields of medicine and prediction and forecasting have led the way in ANN papers. One reason why NN has been so successful in both areas is because the method's structure lends applications, from weather forecasting and agriculture to stock market research. In addition to their usefulness in fields like mathematics and engineering, neural networks may be put to use to assess an organization's needs and then create a customized framework or structure to meet those demands. Research into neural networks also draws from other areas, including business, image processing, neuro-computing, meteorology, and the testing of many types of data. Neural networks, which have recently risen to prominence in the field of image processing, have been extensively included into this investigation. In contrast, image processing emphasizes novel perspectives on data categorization and pattern recognition as essential elements of network design. In image processing, the Self Organizing Mapping (SOM) network is a heritable neural network subtype. Optimizing neural networks using multi-criteria gradients (MCGs) has moved beyond the realm of image processing and into the realms of medicine and forecasting, and will be the focus of future research.

Multiple research have shown that neural networks may benefit from a modified version of the conjugate gradient technique. In order to discover the most precise result, the lowest cost function, and the least error square (MSE), the field of optimization is widely employed, beginning with classical CG and basic Gradient Descend (GD).

While several studies make use of the concrete MCG formulation, others have found a more effective method of enhancing the neural network learning algorithm. A novel finding with improved performance compared to earlier discoveries was achieved by combining different methodologies. As a result, the following table provides a concise summary of recent trends in neural network tweaking that use not just the conjugate gradient method, but also a hybrid learning algorithm drawn from a wide range of optimization methods.

The majority of ANN studies have been on developing more effective learning algorithms for training. In order to accomplish this primary objective, appropriate formulations and algorithms will be used to boost the overall performance of testing data. Otherwise, CG is used as a primary factor in the review's back-propagation error correction. Although extensive study with experimental testing is needed to prove the performance, efficacy, and efficiency of the chosen approach, this is not the case at now. Multiple variants of NN agents and performance factors that might enhance or perhaps be the best answer in that sector can be investigated by expanding the scope of the investigation. Since there are many aspects of neural network learning algorithms to consider, this study may be expanded to include such topics in the future. The network's parameters influencer, the number of neurons and hidden layers, the weight adjustment, and the learning rate may all be tweaked further to obtain the best possible outcome. Therefore, the next step in future research will be to evaluate the new formulation, methodologies, and models with an appropriate data set across a wide range of neural network problem domains.

I would like to express my sincere gratitude to Mr. Prashant Arora, Head of Department Computers and Technology, Welham Boys School for his invaluable support throughout the research.

1. Hinton G E, Srivastava N, Krizhevsky A, et al.(2012) Improving neural networks by preventing co-adaptation of feature detectors. Computer Science, 2012, 3(4): 212-223. 2. C. Szegedy. W. Liu, Y. Jia, et al.(2015) Going Deeper with Convolutions. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. 2015:1-9. 3. Madgwick S O, Harrison A J, Vaidyanathan A.(2011) Estimation of IMU and MARG orientation using a gradient descent algorithm.[C]// IEEE International Conference on Rehabilitation Robotics. IEEE, 2011:1-7 4. Glorot Xavier, A. Bordes and Y. Bengio.(2011) Deep Sparse Rectifier Neural Networks. International Conference on Artificial Intelligence and Statistics.

6. Ke Xu.(2012) Research on application of convolutional neural network in image recognition[D]. Zhejiang University. Proceedings of IEEE. 86(11). 2012:2278-2324. 7. Roza Dastres and Mohsen Soori (2021) ―Artificial Neural Network Systems‖. International Journal of Imaging and Robotics (IJIR), 2021, 21 (2), pp.13-25. ffhal-0334954 8. WANG Xin-gang (2017) ―Overview of the Research Status on Artificial Neural Networks‖ 2nd International Forum on Management, Education and Information Technology Application (IFMEITA 2017) Advances in Social Science, Education and Humanities Research (ASSEHR), volume 130. 9. Enzo Grossi, Massimo Buscema (2008) Introduction to artificial neural network, Lippincott Williams & Wilkins. Unauthorized reproduction of this article is prohibited. 10. Nitin Malik (2005) National Conference on ‗Unearthing Technological Developments & their Transfer for Serving Masses‘ GLA ITM, Mathura, India 17-18 April 2005.

Class 12, Welham Boys School