| Summary: | 碩士 === 國立交通大學 === 電信工程研究所 === 107 === An artificial neural network (ANN) is a computational model that is inspired by the way biological neural networks in the human brain process information. In order to improve not only ANN’s training speed but also the testing accuracy. Inspired by a type of neural glia cells, called astrocyte, have recently been demonstrated to be actively involved in the processing and regulation of the human brain. We propose a novel training algorithm for artificial neuron-glia network (ANGN). The idea of ANGN is contributed by [1] [2]. Through this idea, we applied backpropagation for our training algorithm to compared with Genetic algorithms (GAs) [1] [2]. The accuracy of our ANGN is up to two times greater than [1] [2] if the network is not deep for the Iris flower problem. We also compare our ANN and ANGN with ANN by Keras [3]. Regardless of testing accuracy and training speed, our ANGN outperforms ANN by Keras [3]. Through these advantages, we inspired by [4] and decided to design an end-to-end communications system as an autoencoder for the physical layer. This autoencoder can jointly optimize transmitter and receiver components through training by ANGNs. Our ANGN can completely training by less epochs in the case of uncoded BPSK. Through these advantages, we think that ANGN can effectively improve training process.
|
|---|
