When doing a classification task suppose we have 3 targets with notation -2, -1, 0. I read somewhere it is a good practice to Standardize the labels to positive integers.. In this case suppose we change it to 2, 1, 0. Why is it recommended?
Theoretically I don't think it should matter since each class is unique. But in practice it matters, why?
Please can anybody be kind enough to explain this to me? When doing a classification task suppose we have 3 targets with notation -2, -1, 0. I read somewhere it is a good practice to Standardize the labels to positive integers. In this case suppose we change it to 2, 1, 0. Why is it recommended?
Can you please link a reference to this!. Usually in practice you wouldn't use integer labels at all. Instead you would use one-hot encoded labels. Here since you have 3 output classes, the neural network will have 3 output nodes. To be concrete the classes can be represented as follows :
where one out of the three bits is turned ON. This is because in practice the output of the neural network may be a sigmoid (output: range[0,1]) or softmax function (output: probabilities)
For machine learning frameworks, it is important that the inputs are within well-specified bounds. Requiring target labels to be positive makes subsequent code and interpretation much easier.
When the integers are positive and start from 0, - they can be used as indices to query some collection, like a list of tensors or a list of strings.
