hossein banki-koshki

Neuronal synchronization as a significant cognitive phenomenon of the human brain, has attracted the interest of neuroscience researchers in recent years. This phenomenon is generally investigated in discrete and continuous neuronal models or experimentally recorded signals of the brain. In this study, for the first time, we investigate the weight synchronization instead of neuronal synchrony, in the training step of the artificial feedforward neural networks. The findings of the study show that the generalized weight synchronization occurs both during the training mode and in the non-training mode. Furthermore, as the training is completed, the synchronization increases between the weights. In this study, a new method is introduced in order to detect synchrony patterns using signal derivative and hierarchical clustering. We have also presented a criterion to quantify weight synchronization in different layers of the neural network. Accordingly, the results demonstrate that the lower layers of the network have a significantly higher level of weight synchrony than the upper layers.

The presentation of new neuronal models to simulate cognitive phenomena in the brain has attracted the research interests in recent years. In this study, a new neural model based on the chaotic behavior of weights of artificial neural networks during training by back-propagation algorithm is presented. This model is the first discrete neuronal model with learning ability and shows complex and chaotic behaviors. The learning ability of this model has enabled it to simulate cognitive phenomena such as neuronal synchronization in near-realistic conditions. The model, which is derived from a simple three-layered feed-forward neural network, has several coexisting attractors that make learning possible in various basins of attraction. The study of model parameters shows that bifurcation occurs not only by changing the learning rate, but also external stimulation can change the model behavior and bifurcation pattern. This point that can be used in modeling and designing new therapies for cognitive disorders.

The use of genomic nucleotide sequences as biochemical signals in machine learning methods is possible by converting these sequences into numerical codes. This conversion results in an unrealistic increase in the dimension of the data and encounters some data analysis operations such as visualization and feature extraction with constraints. Therefore, one should use the dimensionality reduction technics in order to return the data to its real dimension. In this study, a deep autoencoder neural network has been used to reduce the dimension of binding site sequence data on the human genome. In order to determine whether the information of real data is preserved in compressed data, we perform a two-class classification using a support vector machine. The results show that information is almost entirely preserved in compression. Then, compressed data is used for visualization as well as feature selection by analysis of variance. The results show that the first, the tenth and eighth positions in the sequences are the most informative positions. While the majority of the previous works deal with gene expression data of microarrays and compare a few dimension reduction algorithms, this paper for the first time uses an autoencoder on nucleotide sequence data and provides a comprehensive comparison between the performance of the dimension reduction technics and machine learning algorithms.