Feed-forward neural network processing speed analysis and an experimental evaluation of Neural Network Frameworks

摘要

Neural Network is an important tool for many pattern recognition, prediction and function approximation tasks. Three Java open source Neural Network Frameworks such as Encog v2.4, Neuroph v2.4 and JOONE 2.0 (Java Object Oriented Neural Engine) are considered here for an experimental evaluation. The performance evaluation is carried out by training the feed-forward neural network to recognize the XOR operator. Implementation of XOR operation is a subset of many complex problems. So it is named as a classic problem in neural network. After training, the output of neural network can be obtained. It may not have the real solution of XOR. However it will be a value so close to the ideal output. To create a benchmark, we developed a sample task. The training technique encompasses an enhanced version of backpropagation which utilizes a momentum to benchmark the neural networks. The backpropagation training method that uses a momentum can yield quick error reduction.