A biologically-inspired network for generic object recognition using CUDA

Abstract

Generic object recognition is one of the most important fields in the artificial intelligence. Some cortex-like networks for generic object recognition are proposed these years. But most of them concentrated on the discussion about the recognition performance (such as recognition rate, number of objects to be recognized), not the practicability, i.e., implementation with ubiquitous devices and application in real time. This paper reports a try on implementation of a biologically-inspired where-what network (WWN), which integrates object recognition and attention in a single network, via parallelizing the various stages of the network training with CUDA on GPU to shorten the training time. The experiment on HAIBAO Robot exhibited in 2010 Shanghai Expo shows that this optimization can achieve a speedup of almost 16 times compared to the C-based program on an Intel Core 2 DUO 3.00 GHZ CPU in real environments. © 2012 Springer-Verlag GmbH.