Is uneven distribution of data across classes in the training sample acceptable?

Question

Neural networks

In the case of training a simple feedforward network with one hidden layer and in the case of pretraining with RBM?

Answer 1

There are 2 cases:
1. In the real world, as in the sample, there is a "skew" distribution by class, that is, some objects are known to be less common
2. In the real world, there is no "skew".
In the first case, it is better to keep the "skew", in the second, generate more data