NOVIDIA, a leading developer of hardware and software used in the development of artificial-intelligence technology, announced that it’s introducing a new series of products and software tools, based on the emerging science of “Artificial Stupidity” (AS). “Artificial Stupidity is not intended to replace AI, but to augment it,” said Jason Wong, founder and CEO of NOVIDIA during a web-based tele-briefing held earlier today.
Wong explained that AS technology, which uses biomimicry algorithms to model the illogical and sometimes counterproductive behavior of humans and some higher primates, is being used by NOVIDIA researchers to soften the hard logic of AI systems. It’s also being considered for the expansion of the optimization criteria used in decision and learning algorithms.
The principles underlying AS theory emerged from an experiment that attempted to solve some of the problems associated with generalized adversarial networks (GANs), i.e. the practice of pitting classical supervised learning AI algorithms against unsupervised learning algorithms.
“In many cases, having the two algorithms compete against each other to solve a problem while sharing their results can lead to more efficient, more complete solutions” said Dr. Aliana Maren, Director of Advance Research at NOVIDIA. “But GANs are very susceptible to erroneous training data which can mislead them and cause serious misclassification events. The math used to create digital models of stupid behavior suggested that adding AS to the mix might improve things, but there was only one way to find out.”
For their experiment, the researchers created a GAN with a supervised algorithm code-named MOE and LARRY, an unsupervised learning algorithm. After a series of exercises that characterized the MOE/LARRY GAN’s performance, they introduced CURLY, an algorithm generated using AS techniques (see figure). The three-way GAN proved to be enormously successful at solving problems, even with flawed or incomplete training data. In these cases, NOVIDIA scientists believe that CURLY’s deliberately induced stupidity pushes the more constrained algorithms to explore possibilities that they would never consider themselves.
Although very promising, some problems still remain. For instance, there’s the tendency for MOE and LARRY to gang up on the unsuspecting third algorithm and replace portions of CURLY’s inference tables with what can only be described as the digital equivalent of slaps, head-butts, and eye-pokes.
Early results indicate that introducing controlled amounts of AS into AI applications can result in new, unexpected behaviors that some scientists have referred to as “Artificial Creativity.” To make her point, Dr. Maren stunned the audience when she told them that “Rails on Fire,” the soul-stirring rock ballad that had topped the pop charts for the past couple of weeks, had been composed and recorded by a competitive chess program they’d been working on, after they’d accidentally exposed it to an AS algorithm.
Other experiments with machine-learning systems have shown that using AS to expand the range and amplitude of the wrong guesses made by an AI system has the potential to greatly accelerate learning times, and make the resulting applications more resistant to problems from erroneous inputs caused by bad data or faulty sensors. Nevertheless, NOVIDIA researchers are proceeding cautiously when applying the new technology to mission-critical applications after an autonomous vehicle using AS algorithms modeled after an undisclosed high-ranking American politician suddenly lit itself on fire and drive off a cliff.
NOVIDIA expects to ship its first artificial stupidity system (ASS) hardware platform and a beta version of its AS software tool kit during the second quarter of 2020. Pricing has not been announced yet.