If a particular system is not trained to perform differently or if it hasn’t been directed to take common variation then it would even fail to identify an image with some stickers.
In the latest turn of events, one of the grad students from the University of Washington, Hossien Hosseini, fiddled around with some colors on different images. As a result, the system failed to recognize the dog and deer which were twisted with a bit yellowish and purple respectively.
Hosseini said in a statement:
“I think we need to find a way for the model to learn the concepts, such as being invariant to color or rotation. That can save the algorithm a lot of training data and is more similar to how humans learn.”
You can have a crack at the complete pre-print paper on Arxiv.
Image Source: Nvidia