This research is to search for alternatives to the resolution of complex medical diagnosis where human knowledge should be apprehended in a general fashion. Successful application examples show that human diagnostic capabilities are significantly worse than the neural diagnostic system. Our research describes a constructive neural network algorithm with backpropagation; offer an approach for the incremental construction of nearminimal neural network architectures for pattern classification. The algorithm starts with minimal number of hidden units in the single hidden layer; additional units are added to the hidden layer one at a time to improve the accuracy of the network and to get an optimal size of a neural network. Our algorithm was tested on several benchmarking classification problems including Cancer1, Heart, and Diabetes with good generalization ability.
Deep Dive into A Constructive Algorithm for Feedforward Neural Networks for Medical Diagnostic Reasoning.
This research is to search for alternatives to the resolution of complex medical diagnosis where human knowledge should be apprehended in a general fashion. Successful application examples show that human diagnostic capabilities are significantly worse than the neural diagnostic system. Our research describes a constructive neural network algorithm with backpropagation; offer an approach for the incremental construction of nearminimal neural network architectures for pattern classification. The algorithm starts with minimal number of hidden units in the single hidden layer; additional units are added to the hidden layer one at a time to improve the accuracy of the network and to get an optimal size of a neural network. Our algorithm was tested on several benchmarking classification problems including Cancer1, Heart, and Diabetes with good generalization ability.
The main characteristic of neural networks (NN) is their ability to generalize information, as well as their tolerance to noise. Therefore, one of the computer science areas that use them the most is Pattern Recognition. The research line of Neural Networks applied to Pattern Recognition has as its main objectives: o The application of Neural Networks to specific problems of pattern recognition. o The evaluation of efficiency metrics and reliability of the solutions proposed. o The formation of human resources in the area. Backpropagation algorithm is the most widely used learning algorithm to train multiplayer feedforward network and applied for applications like character recognition, image processing, pattern classification etc. Before train an artificial neural network (ANN), the network must be built. That is, the nodes in the input layer, output layer and the hidden layer must be defined. The ANNs used for the same problem may differ from each other by their length of hidden layer as the length of hidden layer has to be pre-determined in the traditional backpropagation algorithm. In recent years, many researches have been done on algorithms that dynamically build neural networks for solving pattern classification problems. These algorithms include the dynamic node creation, the cascade correlation algorithm, the self-organizing neural network, and the upstart algorithm. In this research we also proposed an algorithm, which can add nodes in the single hidden layer during the training period and can build an ANN with its minimal size, which can classify patterns with acceptable efficiency.
One of the problems with the traditional backpropagation algorithm is the predetermination of the number of neurons in the hidden layer within a network. To overcome this problem the construction algorithm for feedforward networks may be used, which constructs the network during training. Thus we can have an optimal number of neurons in the hidden layer to attain a satisfactory level of efficiency for a particular problem. Besides applying the early stopping method of training using cross-validation we can also train the network in a relatively short estimation period (training period). In the construction algorithm proposed by Rudy Setiono and Huan Liu they have defined the stopping condition of the training by classifying all the input patterns. It means that while the efficiency is 100%, the training will stop. But in most cases with the benchmarking classification problems 100% efficiency may not be achieved. This is why we used a new algorithm for pattern classification that defines the stopping condition by the acceptance of efficiency level. Another consideration we have made that the desired or acceptable efficiency on the test sets may not be achieved even though the mean square error on training set is minimum. These considerations encouraged us to propose an algorithm that will combine the learning rule of backpropagation algorithm to update weights of the network and the construction algorithm to construct the network dynamically and also consider the efficiency factor as a determinant of the training process.
The following steps are followed to build and train a network [5] ——————–( —————–(2) Where,
Where, k indicates the kth output unit, j indicates the jth hidden unit; i indicates the ith input node, p is the input vector, is the learning rate¡ ¢ is the error term, pi x is the input value to the i,
The error function is usually defined as the meansquared-errors
Where, k denotes kth output unit, n denotes the nth iteration, C is the number of output units, N is the total number of patterns, k d denotes the desired output from k, k y denotes the actual output of neuron k, k e denotes the error term for kth output unit.
We tested our algorithm by three benchmark classification problems Cance1, Heart, Diabetes. The goal is to show that by using minimal size of a neural network we can classify patterns in medical diagnostic data sets with an acceptable efficiency and can also minimize the training period.
As concentrating on the optimal size of an artificial neural network, we have ignored all the additional features (like, error smoothing functions, momentum constant, weight freezing technique etc.) that are used to increase the performance of a backpropagation algorithm, to examine, the level of efficiency of a network with optimal size. Comparing with the previous works on the benchmarking problems we have used and their results, the classification efficiency of our proposed algorithm is quite acceptable. Table 3 shows the comparisons.
We have examined the experimental results by using different learning constant. Hence we can say that for these data sets used, better results are achieved when the £ ¥¤ §¦ ©¥ ¦ ¤ ! ¥" # ¥$ &% § ’ )( !0 1¤ 20 43 5 6 87 @9 BA )C D &E ‘F BA )C G HC I QP ¥R TS UR §V WF YX ¥7 ©X ac onstant used beyond this range did not cause better efficiency.
…(Full text truncated)…
This content is AI-processed based on ArXiv data.
