Artificial Neural Network

Artificial Neural Network Artificial Neutral Network are information processing systems that have characteristics similar to biological neural networks. Artificial Neutral Network are formed as a generalization of the mathematical model of biological neural networks, with the assumption that: 1. Information processing occurs on many simple elements (neurons). 2. Signals are sent between neurons through their connectors. 3. The links between neurons have weights that will strengthen or weaken the signal. To determine the output, each neuron uses an activation function (usually not a linear function) that is applied to the sum of the received inputs. The magnitude of this output is then compared to a threshold. Artificial Neutral Network are characterized by the following 3 things: 1. The pattern of connections between neurons (called the network architecture). 2. The method for determining the connecting weights (called the training/learning method/algorithm). 3. Activation function. For example: neuron Y in the following figure: Y receives inputs from neurons x1, x2, and x3 with relationship weights w1, w2, and w3, respectively. The three neuron signals are then summed: net = x1w1+ x2w2 + x3w3 The amount of signal received by Y follows the activation function y = f(net). If the activation function value is strong enough, the signal will be passed on. The value of the activation function (network model output) can also be used as a basis for changing the robot. History of Artificial Neural Networks Modeling of human biological activity began in the 17th century (study of the working of the heart). Basic studies of neural networks took place in the late 19th and early 20th centuries. Studies were interdisciplinary, involving physics, psychology and neuropsychology. The scientists involved were Hermann von Helmhotz, Ernst Mach, and Ivan Pavlov. The modern view of neural networks emerged in the 1940s through the work of Warren McCulloch and Walter Pitts, who showed that artificial networks of neurons could, in principle, compute all arithmetic and logical functions. Donald Hebb followed by proposing the concept of learning mechanisms in biological neurons. The first practical ARTIFICIAL NEUTRAL NETWORK applications appeared in the late 1950s, with Frank Rosenblatt's invention of perceptron networks and associated learning rules. In the same year, Bernard Widrow and Ted Hoff introduced a new learning algorithm, which is still in use today. In 1969, Marvin Minsky and Seymour Papert pointed out some inherent limitations of Rosenblatt and Widrow's network. At the time, some people believed that ARTIFICIAL NEUTRAL NETWORK research was over. The doubts at that time (the 60s) arose due to the lack of new ideas and compu performance. ARTIFICIAL NEUTRAL NETWORK research continued in the 70s, when Teuvo Kohonen and James Anderson separately developed new ARTIFICIAL NEUTRAL NETWORKs that could function as memory. The 80s saw the advent of high-performance computers (for the time) and ARTIFICIAL NEUTRAL NETWORK research increased dramatically. Stephen Grossberg (1976) was also very active in researching self-organizing networks. John Hopfield (1982) introduced the concept of statistical mechanics, which was used to explain recurrent networks, which could function as memory. The second key advance in the 1980s was the invention of the backpropagation algorithm to train multi-layer perceptrons by a number of researchers. The most influential publication of the algorithm was written by David Rumelhart and James McClelland. This algorithm was an answer to Minsky and Papert's (1960) criticism. Research continued. New ARTIFICIAL NEUTRAL NETWORK concepts, algorithms, and architectures emerged. What will happen 20 years later? Artificial Neural Network Application In 1988 the DARPA Neural Network Study listed various applications of ARTIFICIAL NEUTRAL NETWORKs, beginning with the application of the adaptive channel equalizer (1984), which was a single-neuron network for long-distance telephone systems, to stabilize voice signals. This tool experienced tremendous success in the industry. In addition, some applications of ARTIFICIAL NEUTRAL NETWORKs: 1. Pattern recognition ARTIFICIAL NEUTRAL NETWORKs can be used to recognize patterns (e.g. numbers, letters, sounds, or signatures) that have changed slightly. This is similar to the human brain which is still able to recognize people it has not seen for some time. ARTIFICIAL NEUTRAL NETWORK signal processing (ADALINE model) can be used to determine noise in telephone lines. 2. Forecasting ARTIFICIAL NEUTRAL NETWORK can also be used to predict what will happen in the future based on past event patterns. This can be done given the ability of the ARTIFICIAL NEUTRAL NETWORK to remember and make generalizations from what has come before. 3. Aerospace High accuracy aircraft autopilot, flight path simulation, aircraft control system, aircraft component simulation, aircraft component damage detector. 4. Automotive Automatic guidance system, driver activity analyzer. 5. Banking Check and document reader, credit application evaluation 6. Defense Weapon control, target tracking, object discrimination, face recognition, new sensors, sonar, radar and image signal processing (including data compression), feature extraction, noise suppression, signal/image identification. 7. Electronics Code sequence prediction, integrated circuit layout, process control, chip defect analysis, machine vision, sound synthesis, nonlinear modeling. 8. Entertainment Animation, special effects, market forecasting 9. Finance Property valuation, credit advisory, mortgage examination, corporate bond rating, credit analysis, portfolio sales program, financial analysis, exchange rate prediction. 10. Insurance Policy application evaluation, production optimization 11. Manufacturing Process control, product analysis and design, machine and process diagnosis, particle identification (real time), quality inspection system (visual), packaging, quality analysis, etc. 12. Medical Cancer cell analysis, EEG and ECG analysis, prosthesis design, transplant time optimization, hospital expenditure reduction, hospital quality improvement. 13. Oil and gas Exploration 14. Robotics Trajectory control, robot lifting, vision system 15. Conversation Conversation recognition, compression, letter classification, text-to-speech synthesis. 16. Securities Market analysis, automatic bond rating, stock trade reporting system 17. Telecommunications Image and data compression, automated information services, automated translators, account payment processing systems 18. Transportation Brake diagnosis system, scheduling, routing Limitations of Artificial Neural Networks Although many applications can be performed by the ARTIFICIAL NEUTRAL NETWORK, it has some general limitations. The main limitation is the "INACURACY" of the results obtained (because the ARTIFICIAL NEUTRAL NETWORK works based on the patterns formed on its inputs).