ALT Image showing how if we have a softmax classifier with 4 classes but only 2 features, there exist choices of softmax weights for which one of the classes cannot be assigned the largest probability. The image on the left shows a 2D plane partitioned into 3 cones, which correspond to regions where the probabilities for class 1, class 2 and class 3 are maximised. There is no region where class 4 has the largest probability. On the right are four corresponding subplots, one per class, showing the softmax probabilities assigned to each classes as we move in feature space. The plot for class 4 looks like a small hill that does not surpass a certain probability threshold, while for the other 3 classes it looks like a wave that reaches a probability of one when moving outwards in a particular direction of feature space.