The Next Top Model

See allHide authors and affiliations

Science  01 Jan 2010:
Vol. 327, Issue 5961, pp. 13
DOI: 10.1126/science.327.5961.13-d

Consumers may be familiar with high-end graphic processing components in video game consoles, such as the PlayStation3, or as a consequence of outfitting personal computers ordered online with NVIDIA graphics cards; these advances in hardware have also attracted the attention of procurement officials in the military services. In the academic realm, Pinto et al. have harnessed the power of clustered graphics processors to assess the relative performance of machine vision models of object recognition. The availability of massively parallel processing power at reasonable cost allowed them to explore, in 1 week versus 2 years, sizable regions of parameter space by varying the number of filters, the learning rate, and so forth. They generated a library of 7500 models that were trained on individually rendered objects during an unsupervised learning phase, and then screened on the basis of recognizing cars versus planes, which were presented in a range of orientations and on a variety of backgrounds. The top-ranked models were then evaluated broadly across other objects and on one of the toughest recognition tasks—photographs of human faces—and compared to a number of sophisticated algorithms, which yielded a small set of parameter values that were associated with high object recognition accuracy.

PLoS Comput. Biol. 5, e1000579 (2009).

Navigate This Article