Teaching network connectivity using simulated annealing on a massively parallel processor
Article Abstract:
Simulated annealing on a massively parallel processor can be used to teach network connectivity. Finding an optimum connectivity pattern is the object of training a network. There are two layers in the network. The first layer receives an 8-but gray level image and provides four separate directions to supply local edge detection. The first layer and the second layer have a sparse connectivity. An indication of the presence and location of trained objects is received from the second layer's output. The energy of the system is lowered gradually during annealing, requiring that the energy of a particular connection scheme be defined so that annealing permits quick convergence to a good set of connections. A relationship between the neural model and a physical thermodynamic model is needed to provide a method to train connections using simulated annealing.
Publication Name: Proceedings of the IEEE
Subject: Electronics
ISSN: 0018-9219
Year: 1991
User Contributions:
Comment about this article or add new information about this topic:
Massively parallel computational simulations in light scattering
Article Abstract:
Explaining natural meteorological phenomena, the color of the sky, for example, or the physical characteristics of individual particles and particular systems requires simulating the scattering of light by small particles. Particle size and concentration can be determined by using light scattering instruments. A geometrical optics approach can determine light scattering when the object is much larger than the incident light's wavelength. Parallel calculations can be done on how individual rays scatter light. The Connection Machine, a finely grained, loosely coupled SIMD machine, can be used to study light scattering. Since this is an SIMD machine, the starting parameters of a processor can be reset so the processor can be assigned a new ray as soon as it finishes with a ray, resulting in improved processor efficiency.
Publication Name: Proceedings of the IEEE
Subject: Electronics
ISSN: 0018-9219
Year: 1991
User Contributions:
Comment about this article or add new information about this topic: