While doing logistic regression, it is common practice to use one hot vectors as desired result. So, no of classes = no of nodes in output layer. We don't use index of word in vocabulary(or a class number in general) because that may falsely indicate closeness of two classes. But why can't we use binary numbers instead of one-hot vectors?
i.e if there are 4 classes, we can represent each class as 00,01,10,11 resulting in log(no of classes) nodes in output layer.
It is fine if you encode with binary. But you probably need to add another layer (or a filter) depending on your task and model. Because your encoding now implicates invalid shared features due to the binary representation.
For example, a binary encoding for input (x = [x1, x2]):
'apple' = [0, 0]
'orange' = [0, 1]
'table' = [1, 0]
'chair' = [1, 1]
It means that orange and chair share same feature x2. Now with predictions for two classes y:
'fruit' = 0
'furniture' = 1
And linear optimization model (W = [w1, w2] and bias b) for labeled data sample:
(argmin W) Loss = y - (w1 * x1 + w2 * x2 + b)
Whenever you update w2 weights for chair as furniture you get an undesirable update as if predicting orange as furniture as well.
In this particular case, if you add another layer U = [u1, u2], you can probably solve this issue:
(argmin U,W) Loss = y - (u1 * (w1 * x1 + w2 * x2 + b) +
u2 * (w1 * x1 + w2 * x2 + b) +
b2)
Ok, why not avoid this miss representation, by using one-hot encoding. :)