Collage: Karen Zack/@teenybiscuit


As far as possible, artificial intelligence should eliminate human error. However, even an AI application is not totally error-free. Image recognition systems are particularly prone to making mistakes. Ulti-mately, the neural network behind image recognition is only ever as good as its training data.
The fact that these algorithms are easily tricked is demonstrated by the internet meme Chihuahua or Muffin? – similar-looking content that is instantly recognisable for any human being still causes major problems for most software.
These mistakes are often quite funny. Yet some can have serious consequences – applications for autonomous vehicles that misjudge a traffic situation, for example. It only takes a sticker or some graffiti on a traffic sign to make accurate image recognition difficult.
If we are to realise the dream of a technology that is equal or even superior to humans in terms of thinking, these errors must be eliminated.

With thanks to Horst Bischof at the Institute of Computer Graphics and Vision/Graz University of Technology.