Creating a Neural Network Using EasyNN-Plus - Essay Example

Creating A Neural Network Using EasyNN-Plus A Neural Network is defined as an interconnected assembly of simple processing elements, units or nodes, whose functionality is loosely based on the animal brain. The processing ability of the network is stored in the inter-unit connection strengths, or weights, obtained by a process of adaptation to, or learning from, a set of training patterns. Neural nets are used in bioinformatics to map data and make predictions. As EasyNN-Plus is tool/program/software used to create a neural network system for Microsoft Windows, it helps the creation of neural networks easy. It allows the user to produce multilayer neural networks from a grid or from text files and images. The user can produce training, validating and querying files using the facilities in EasyNN or using any editor, word processor or spreadsheet that supports text files. EasyNN can learn from training data and can self authenticate while learning. It can be queried from a file or interactively. EasyNN can produce spreadsheet like output and results files. All graphs and diagrams are restructured during training and querying so the user can see how the neural networks are working. The EasyNN-Plus has a number of shortcuts, and power keys which allows an advance user to carry out their task quiet efficiently. Despite the fact that for the new users who are not conversant in it can find it quiet difficult in the beginning and puzzling as to how to use the software, and this may take a lot of time and patience. At the same time as using EasyNN the user does not in fact learn how to create the neural network, as EasyNN mechanize the process of producing a neural network the steps produce the network is quiet unseen from the user. EasyNN uses mathematical literacy as the backbone of the information provided by the user. This lets EasyNN-Plus to provide several views of a neural network, and also increases the uses of using EasyNN-Plus to produce neural networks for a few different things e.g. forecasting, data validation, and customer research. EasyNN used the algorithm as back propagation. The differences from most other back propagation based applications are the data structures and the way the data is presented to the learning algorithm. EasyNN uses double linked lists to store the examples, the nodes and the connections. The lists can be managed quickly in both directions at the same time. The lists can also be extended and contracted dynamically. The EasyNN estimates the number of neurons required and determines how the network will learn. It maximizes the learning rate and thrust by running a few learning cycles with different values before actual training. EasyNN can then automatically decrease them if there is inconsistent learning or if oscillations in training error take place. The user can change these values, however again to be able to do so with confidence requires considerable knowledge about the software, the data set and the neural network. A specific advantage of using EasyNN software is that it has validation built in. This is used to avoid the neural network from becoming "overfitted" to the training data. The user can select the percentage of the training data to be used for validation. Another useful tool in EasyNN is the integrated function for handling missing values. This function replaces the missing value with the median of the affected attribute in the training dataset in both training and evaluation. The data being used here is medical information on possible patients suffering from diabetes. A number of patients may or may not have diabetes and this network is going to be built to predict whether or not a patient has diabetes with the data available. The first stage in building a network using EasyNN is to provide the software with the appropriate data it is going to need. In this instant a .txt file was imported into EasyNN containing all the data the program require to build the network including the names for each column and their values. Other formats containing data may also be used such as CSV or XSL files to build a network in EasyNN. EasyNN means to be told how to separate one column with data from another and it provides the user with the choice to choose how the data is separated from one another in the original file so that it makes out the different column heading, and in this instance a "," (comma) was used, then EasyNN was programmed so it recognised what numeric value each column held such as 'Integer' or 'Real'. When the data has been imported and established as to what value it hold the network can be built and it can start learning by performing cycles to disseminate errors from the network, and then tasks such as querying and simplifying the network can be carried out. The network learns and the user sets the error level as to which the network will output the specific value (average). The network also requires to be told how many cycles to go through in order for it to grow and learn. As soon as EasyNN has built a network, there are several visual representations available to the user that displays what has been created and in how many cycles etc. This lets the user to study the network and see if its performing correctly. Example of test data Network The above graph represents how applicable the network finds each column to that task that it is trying to carry out. If a specific column is not applicable it maybe slowing the network down and changing the outcome of the data as the weights of the input and out may be affected. The manager may use this representation to help them 'simplify' the network hence preventing the network from overfitting. A number of the columns were removed from the network and some of the columns had there element changed to output, so that one could see what the network was expecting and then comparing it to the original data. Moreover only some of the rows were used to establish the effectiveness of the network and the learning process. On the whole all EasyNN is quiet an easy and motivating program to use. It is quiet supportive and easy to use for users who do not have much knowledge on neural networks. Users may find it hard to understand and create the network initially however as they use it more often, the program tends to become easier to understand and use. The program automatically calculates the weights and hidden layers that the network requires so to predict conclusions of the data that has been presented to the program. Simplifying the Network The network will now be simplified to try enabling it to run faster and more precisely by removing some fields that are not relative and to do this the graph containing the relevance of each field will be used and the column with the least relevance will be removed from the network building process After studying the graph, the pregnancy column does not seem to have much relevance to the network so now that will be removed. Theoretically the envisaged data should get more precise and the network should get smaller and run faster. Test 1 After Pregnancy was removed. Overview After pregnancy was removed, the importance of some of the inputs had Test 2 After ID was removed Taken as a whole the theory of the network learning faster and being more accurate was correct, however so as to test the network, the question arises what would happen if one of the most relevant inputs to the network were removed, to find out in the next test the glucose column will be removed. Test 3 after Glucose removed After the glucose was removed the network relied on the information on in the body mass column thus making more relevant to the network. As one can see the error line is not as straight as the ones in the other, as this data was for predicting whether or not patients have diabetes, glucose levels is one of the main factors in predicting if someone has diabetes or not as a result the network is producing a lot more errors now that the glucose input has been removed. EasyNN has the ability to allow the users to establish how soon the program may stop learning; one can not in fact assess how long the training of the actual system may take. It depends on the number of inputs, the weight of the inputs and so the number of cycles and learning times may increase or decrease, and if the training is stop precipitately then this may have an impact on the predict results that the network may produce as it has not had enough time to learn. Not all of the predictions when matched up to the information available to the supervisor were correct and precise although in the majority of the cases the predictions were spot on. On the whole the network can be simplified and trained and programmed to learn quicker however the relevant information needs to be entered in the network in order for it to achieve this. The information used was information on whether a patient was suffering from diabetes using there medical information such as glucose, age, blood pressure etc. An Assessment Of Rule-based & Connectionist Neural Networks Neural networks have been intended to create and create the thinking of a human brain. The human brain has progressed and developed over a number of thousands of years, and also has the capacity to adjust to any situation and change the brains way of think, whereas neural networks and computers are very inactive. Despite the fact that they may be able to adjust to certain/rare situations they mainly rely on rules usually programmed in by a human. (Internet, 2006) Despite the fact that rules based systems try to simulate a human expert, the knowledge contained in the brain of the human expert is far more advanced to that in a rule-based system. The rule-based system is limited to the rules and the construction that were originally programmed in to the system. Consequently the original rules and architecture of the rule-based system need to be accurate, and if this is not correct then the conclusions provided by the rule-based system will be worthless and imprecise. At the same time as artificial intelligent systems such as rule based and neural networks need to be trained while being supervised, this perhaps guidance in setting up the rules in a rule based system or sample data in neural networks. Human brains have the ability to learn unsupervised and without guidance, making them very capable. In addition Brains are able to accept ambiguity, to learn on the basis of very poor and disorganised data; at the same time as much more accurately formatted and structured data and rules are required by intelligent systems. Neural networks learn to carry out tasks and study data rather than directly being programmed to do so. Neural networks have the ability to 'learn' how to perform a task and so do not have to be directly programmed to carry out a task, thus removing the presence of a human expert. Neural networks take into consideration and concern the different factors that may affect the particular task and these factors can be used to make predictions by the network. Just like a human brain may allow for different factors for the surrounding environment. On the other hand, neural networks and expert systems will always have there advantages over human experts, for instance; 1. Intelligent systems do not have to be paid, 2. They do not need holidays, 3. Do not get ill, 4. Do not have personal problems to deal with 5. Do not need breaks An intelligent system knowledge can be transferred very easily from one computer to another, and is permanent in carrying out the task is programmed to do. Moreover the information can be stored permanently. On a worldwide basis the expert systems would be of very use to organisation as if there is a human expert and he is needed in a another branch in another country, then it would take a lot of time and money for that organisation to try and get that expert there, at the same time as if there was an implemented expert system the software could be quiet easily made available via tools such as the internet. However the one thing that expert systems may never be able to match up to humans is the fact that human experts may have an insight in specific problems and have the ability to adapt at the split of a second and use 'judgement' in certain situations whilst expert systems don't have this facility as yet. The Function of Intelligent Systems within an Organisation Industrial, commercial and financial fields have had much use for intelligent systems in their organisations. It has facilitated organisations to improve their capability and resourcefulness in a Knowledge-Driven Economy, and has already been applied in many industries. Expert systems in organisations fall into a number of different categories and have been rather successful where human decision support is involved. The applications available to organisations that help in human support decisions fall in to these categories. 1. Diagnostic System 2. Planning & Scheduling systems 3. Interpretation systems 4. Prediction systems These categories show that, expert systems are not just being used to calculate tendencies or forecasting but also being used to solve quite complex and difficult problems in particular fields. When an organisation decides to put into operation an expert system in a particular field or to do a job that a human usually has control over and is dependant on human knowledge, the organisations will have to take in to account various different issues before they apply the system. After the organisation has taken these issues and all the issues related to creating an expert system, and considered the cost of programming and customising such a system. They need to start thinking about whether such a system would be advantageous for the organisation, and it may not seem so in the short run but may turn out to be quiet an efficient and useful system in the long run. In general expert systems have major roles to play in organisations and there have been several successful applications. Organisations and societies altogether have a lot to gain from expert and intelligent systems that can think and understand and make decisions for themselves. Once put into operation the systems are cheaper and more efficient than human in the particular field they have been trained to carry out their tasks. This could make it very attractive for an organisation as whether they actually employ humans to do the task or computers, thus making it a principled issue. However in some cases humans are just not as quick and efficient as computers such as mathematical situations, as a result making it more viable for organisations to have computers rather than humans. BIBLIOGRAPHY Artificial Neural Networks in Medicine[Internet]Available at: http://www.emsl.pnl.gov:2080/docs/cie/techbrief/NN.techbrief.ht [Accessed 01/03/2006] Buchanan Bruce G. and Shortliffe Edward H., editors, (1984). Rule-Based Expert Systems --The MYCIN Experiments of the Stanford Heuristic Programming Project, Reading, MA: Addison-Wesley Barnard, E., and Cole, R. 1989. A neural net training program based on conjugate gradient optimization. Oregon Graduate Centre Tech. Rep. No. CSE 89--014, Oregon Carpenter, G., and S. Grossberg. 1988. The ART of Adaptive Pattern Recognition by a Self-Organizing Neural Network. IEEE Computer 21 Connectionist Expert Systems {1994}[Internet] Available at: http://heaven.eee.metu.edu.tr/vision/books/alife/ch3.html [Accessed 20/12/2005] Diagnosing Why a Car Won't Start - online demonstration of an expert system. Available at: http://www.expertise2go.com/webesie/car/ EasyNN Plus Frequently Asked Questions [Internet] Available at: http://www.easynn.com/ENNPlusFAQ.htm[Accessed 30/01/2006] EasyNN-Plus Help Manual, (2003). Neural Planner Software Elliot, J (2005-2006) Intelligent Systems Lecture notes, Leeds Metropolitan University Luger George, (2005). Artificial Intelligence: Structures and Strategies for Complex Problem Solving, 5th Edition, Addison-Wesley Lawrence, J. Introduction to Neural Networks and Expert Systems. California Scientific Software, Nevada City, CA, 1992. Methods of rule based systems {1994}[Internet] Available at: http://ai-depot.com/Tutorial/RuleBased-Methods.html[Accessed 22/01/2006] Neural Networks [Internet] Available at: http://www.doc.ic.ac.uk/nd/surprise_96/journal/vol4/cs11/report.html [Accessed 02/03/2006] Neural Networks: An Overview[Internet] Available at: http://www.ccs.neu.edu/groups/honors-program/freshsem/19951996/cloder/[Accessed 22/01/2006] Notes on Neural Networks learning and training[Internet] Available at: http://www.generation5.org/content/2004/NNTrLr.asp[Accessed 25/01/2006] Special Interest groups on Artificial [Internet] Available at: http://www.igda.org/ai/report2003/aiisc_rule_based_systems_report_2003.html[Accessed 17/03/2006] [Internet] Available at: http:// www.krl.caltech.edu/charles/alife-game/glossary.html[Accessed 12/12/2005] Read More