AI does it wrong! For those who work in IT one of the big challenges of AI is to learn how to cope with AI mistakes.
When we write code in a deterministic environment we have mastered testing and debugging techniques that help us fix problems when they arise.
Most of the time, if the complexity of the system we are working on is not too big, we are able to clearly understand what caused “the bug” and find a remedy.
AI is a little bit different. When we provide to an AI system the photo of a cat and we ask to confirm that is indeed a cat we can be mesmerised by our system telling us “it thinks” the photo is that of an ambulance.
How can it be so wrong?
The short answer is: is a vector problem.
Imagine a vector as a finger pointing to something.
For some strange reasons our system, receiving a photo of a cat, should point his finger to the cats domain and instead is now pointing its “virtual finger” to the domain of vehicles.
This has probably something to do with the quality and number of cat’s photos we have used to train our system to distinguish the animal from a car.
Improving our training set can help sort out the issue.
Machines have a different way to look at things so, elements in a picture that our brain is filtering, can trick the system causing a change in the direction our virtual finger is pointing to.
So changing elements in the background of the image let the human eye still see a cat, while our virtual finger can move from the cats’ domain into the vehicle domain.
These same criteria apply to natural language understanding hence the reason why the distance between domains can play such a crucial role.
AI does it wrong!Fortunately for us, building bots, some level of misunderstanding can help make our “machines” look somewhat more “human”.