3.4 Understanding the problem with typical neural networks
The deep neural networks we discussed in previous sections are extremely powerful and, paired with appropriate training data, have enabled big strides in machine perception. In machine vision, convolutional neural networks enable us to classify images, locate objects in images, segment images into different segments or instances, and even to generate entirely novel images. In natural language processing, recurrent neural networks and transformers have allowed us to classify text, to recognize speech, to generate novel text or, as reviewed previously, to translate between two different languages.
However, these standard types of neural network models also have several limitations. In this section, we will explore some of these limitations. We will look at the following:
How the prediction scores of such neural network models can be overconfident
How such models can produce very confident predictions on OOD data
How tiny, imperceptible...