Weird Deep Neural Networks

A collection of examples of weird inadequacies in the object recognition performance of deep neural networks.

Two otter three ‘aint bad


The InceptionResnetV2 CNN is pretty sure the left and right images are otters, P(Otter) = .8, but is not at all confident about the middle image, P(Otter) = .1. Azulay & Weiss (2019, Figure 1).