I have 20 years meteorological data of a weather station. Some of the parameters are negative values. Before we can train ANN with this data set, how can these negative values be treated? Do we need to apply data transformation/normalization or in the raw form we can train ANN?
It is always better to normalize the data before giving as training input. There are several methods of this normalization. I think min-max normalization will work for you.
It depends on the type of neurons you are using.
Normally, weights are floating point values that can adapt
to any value between -infinity and + infinity. If the input of a
neuron is negative and the activation should be positive,
the neuron will develop a negative weight so that the product
of input times weight will assume a positive value.
For example, in the case of a sigmoid function, negative
activation is allowed.
http://www.wolframalpha.com/input/?i=1%2F%281%2Bexp%28-x%29%29+for+x+%3D+-20+to+20
Of course, as Manasi Patil said, you should normalize
first. To avoid the 'saturated' ranges, where a slight
change in input does not lead to a considerable change
in output.
Regards,
Joachim
- Badges
- Science topic
- Similar topics
- Computer Communications (Networks)
- Simulators