Investigating Architectural Issues in Neuromorphic Computing.

摘要

Our objective is the study of an Intelligent Text Recognition System (ITRS) that mimics the human information processing procedure to fill in the missing or damaged text by considering the word level and sentence level context. The system is built upon two cognitive computing models, the Brain- State-in-a-Box (BSB) Attractor Model and the Cogent Confabulation Model. The former performs character detection and later performs word and sentence completion. Given a scanned text image where each character is 15-by-15 pixels large, experimental results show that, when 20% of the character images are damaged by a 1-pixel-wide horizontal scratch running through the center of the image where most of the information to distinguish amongst various characters is found, the ITRS recognizes complete sentences at 92% accuracy. When 60% of the character images are damaged by a 3-pixel-wide horizontal scratch located at the center, the ITRS recognizes sentences at 64% accuracy.