Researchers Highlight AI Vision Weaknesses
The technology behind it may well be impressive as are some of the things it can do, but artificial intelligence can also make some pretty embarrassing errors.
AI vision has a weak spot as has been pointed out by researchers from UC Berkeley, and the Universities of Washington and Chicago.
They've drawn up a list of 7,500 images that artificial intelligence systems are struggling to identify correctly.
Researchers tested several machine vision systems on this dataset and discovered that their accuracy dropped by as much as 90 percent. In some cases, the software was only able to identify two to three percent of the images.
Among the catalog of errors that AI has made has been to mistake butterflies for washing machines and dragonflies as bananas.
While this is all slightly comical there are some serious concerns, especially as machine vision systems are going to be placed at the heart of such new technologies as self-driving cars and AI security cameras.
Improving Accuracy
To accompany the dataset the researchers have released a paper - Natural Adversarial Examples - that covers their work and explains the reasons for the flaws, which include AI's "over-reliance on color, texture and background cues" to identify what it sees.
In compiling a long list of AI mistakes the researchers are demonstrating how vision systems can make unenforced errors. They want to draw attention to the issue and hope that the database will help others to improve the accuracy of how artificial intelligence systems classify images.
Add your Comment
[LOGIN FIRST] if you're already a member.fields are required.