The application of noisy reward/penalty learning to pyramidal pRAM structures

It is shown that the addition of noise during probabilistic RAM (pRAM) training develops the property of generalization and therefore the ability to recognize patterns in noisy images. Global reward/penalty learning applied to the pRAM was shown to be an efficient training method that was also hardware-reliable. Results are presented for a pRAM net which show that successful discrimination of patterns can be achieved in the presence of over 45% noise, with a 20% confidence margin.<>