A NEURAL NET FOR 2D-SLOPE AND SINUSOIDAL SHAPE DETECTION

Abstract

2D-slope and sinusoidal shape detection are application specific tasks which are widely discussed in the literature. A neural network is presented which is able to learn a set of different slopes or a set of sinusoids of different frequencies and to detect test patterns after the training stage. The neural net is composed of input neurons, delay neurons and output neurons. The delay neurons form a set of tapped delay lines. Each delay line adapts to its specific signal propagation velocity. The signal propagation velocity vector field of the delay lines is learned by collectively tuning the signal propagation velocities. The neural net is fed with a set of spatiotemporal training patterns, such as bars of different slopes or sinusoids of different frequencies. After training, the net is tested with a random set of 2D- patterns. Unsupervised learning with a Boltzmann temperature term is assumed.