The Failure Modes of Machine Learning

Machine learning systems are brittle. You can run into unfortunate situations like the following. A problem there is that horses tend to be on greener backgrounds because they're around grass. And then what happens when you have a cat in the desert? So this is a well known failure mode of these neural networks. They can just be detecting spurious correlations. It could still handle a terrier, at least distinguish them from other mammals or something. The nice thing about language is that unlike images, images are much more noisy.