An evaluation of a self-supervised topographic neural network using correlations.

摘要

We evaluated Luttrell's self-supervised topographic neural network which is an unsupervised and self-organizing network. We compared its performance to that of other self-organizing neural networks including Kohonen's self-organizing feature map. In a simulation study, we used the network as a two-channel vector quantizer to reconstruct two correlated vectors and used the reconstruction error as a measure of its ability to use the correlation between different channels. In another study, we devised a unique scheme for power load forecasting using Luttrell's network, and we used actual power system data to compare its performance with that of conventional algorithms.; In our simulation studies, we used two mathematical models to provide data to two channels of a self-organizing network. In the first, we used pairs of two-dimensional vectors where the second vector was a rotated version of the first with a random angle of rotation whose statistics defined the degree of correlation between the two vectors. The second (referred to as the AR-like model) also involved vector pairs where the second vector was a randomly perturbed version of the first. The size of the random perturbation, defined by a coefficient in the AR-like model, controlled the correlation between the two vectors.; The simulation results indicated Luttrell's network had a lower reconstruction error than other self-organizing networks. Also they showed that the mean rotation angle of the rotation model and the coefficient in the AR-like model were related to the reconstruction error with Luttrell's network but not with other self-organizing networks. These findings indicated that Luttrell's algorithm took advantage of the correlation in the pair of patterns processed by the two channels.; Using Luttrell's unsupervised network, we devised a compact, real-time adaptive power load forecaster with better performance than conventional load forecasters. We obtained less than a 1% error and around a 2% error in hour-ahead and day-ahead forecastings, respectively. The network could be trained in a few minutes using the most recent historical data.; Our work shows the advantages of the self-supervised network. These encouraging results should stimulate others to explore its use in place of supervised networks in other applications.