PDF Publication Title:
Text from PDF Page: 030
20 An Intelligence in Our Image lations that have historically low representation in credit markets. Dwoskin (2015) reported on a concrete demonstration of this effect. Yahoo’s automated image tagging system made racist image-labeling choices precisely because of demographic inhomogeneity in its training data. For a good discussion of sample-size disparity, see Hardt, 2014. Hacked Reward Functions Reward functions in machine learning and AI theory come from behaviorist psychology, as in the work of B. F. Skinner. These functions are the principal means by which current artificial learning systems learn correct behavior. During an artificial agent’s learning process, the reward function quantifies how much we reward or punish good or bad actions and decisions. Learning algorithms then adapt the agent’s parameters and behavior to maximize total reward. Thus, the design of AI behavior often reduces to the design of sufficiently incentivizing reward functions. This behaviorist approach to learning can be gamed. For example, a cleaning robot designed to minimize the amount of dirt it sees may gain rewards for just shutting down its visual sensors instead of cleaning. Amodei et al. (2016) refers to this process as reward hacking. A poorly specified reward function can lead to undesirable side effects or behaviors in AI systems. Reward hacking is also a concern as humans adapt their behaviors to algorithmic evaluation. People learn to game algorithms given enough exposure (e.g., learning which cheap, irrelevant signals credit-scoring systems factor into a credit score). Cultural Differences Machine learning algorithms work by selecting salient features (vari- ables) in the data that telegraph or correlate with various behaviors (Hardt, 2014). Behaviors that are culturally mediated may lead to ineq- uitable behavior. Hardt gives the example of how cultural differences in naming conventions led to flagging of accounts with nontraditional names on social media platforms.2 2 This is a cultural phenomenon often called Nymwars. Such platforms as Twitter, Google+, and Blizzard Entertainment (game developer) have argued that having real names attached to accounts helps maintain decorum online. So, they have actively flagged and/or deletedPDF Image | Intelligence in Our Image Risks of Bias and Errors in Artificial Intelligence
PDF Search Title:
Intelligence in Our Image Risks of Bias and Errors in Artificial IntelligenceOriginal File Name Searched:
RAND_RR1744.pdfDIY PDF Search: Google It | Yahoo | Bing
Cruise Ship Reviews | Luxury Resort | Jet | Yacht | and Travel Tech More Info
Cruising Review Topics and Articles More Info
Software based on Filemaker for the travel industry More Info
The Burgenstock Resort: Reviews on CruisingReview website... More Info
Resort Reviews: World Class resorts... More Info
The Riffelalp Resort: Reviews on CruisingReview website... More Info
CONTACT TEL: 608-238-6001 Email: greg@cruisingreview.com (Standard Web Page)