Welcome Stranger to OCC!Login | Register

Teaching AI with Partial Credit

Category: Science & Technology
Posted: 02:26PM

According to science fiction artificial intelligence will one day lead to machines that can at least emulate emotions, if not actually experience them. We are not there yet, and likely will not be for a long time as we are still trying to teach systems how to identify various objects. Researchers at MIT have decided to try a new approach for training machine learning systems by awarding partial credit for nearly right identifications.

One use of machine learning systems is for scanning images and identifying objects and actions in the scene. Traditionally if a systems gets it wrong, it is told it is and then moves on. What the MIT researchers are doing is awarding partial credit to the system if its wrong answer was close enough to the right one. If instead of tagging an image with 'sunshine' it tagged with 'summer,' the system would be given some credit because sunshine and summer are likely to co-occur. Traditional methods would score the system as wrong as if it had tagged the image with 'rhinoceros' in place of 'sunshine.' The images and human-sourced tags were gathered from Flickr.

To make it possible to award an appropriate score, the researchers turned to the Wasserstein distance, which compares probability distributions and thus provides a metric of how correct the machine learning system was. Beyond helping the AI learn, this approach could also make an AI more useful for people, as sometimes the terms we use to search for something are not exactly correct, but still similar.

Source: MIT

Register as a member to subscribe comments.

This news has comment postings disabled because it is now archived.

© 2001-2018 Overclockers Club ® Privacy Policy
Elapsed: 0.3351149559   (xlweb1)