New to MemBrain, and I like what I see !
While playing with the Lesson Editor, I saw that ONLY numerics are imported ?! So a date like 08/12/2011 is imported as 08, value 'Oct' is imported as 0. Do I really have to decode all to numerics before I can import raw CSV ? Hope that someone can help ...
RAW CSV Import: why ignore DATES or CHARACTERS ?
Re: RAW CSV Import: why ignore DATES or CHARACTERS ?
Hello and welcome!
What would you expect a date/time information to be converted during import or how should the NN process this information, respectively?
Additionally, date/time information most of the time is not a good type of information to be provided to an NN's inputs anyway, even if converted to a numeric value. This is due to the fact that date/time information by nature always evolves towards higher values and thus does not represent an information that can be used for generalization or for learning association rules: The date/time information by definition always differs from data set to data set.
The only way to incorporate date/time information in a possibly useful way is to convert it to some periodic value that represents a period which might be meaningful for the problem: E.g. a period of a year can be useful information for some problems that might be affected by time of year. In such a situation date/time information may be converted into a sine wave signal in periods of one year. This allows the NN to learn rules that are related to the time of year. Other problems may get along better with a time-of-day periodic signal.
What is your date/time information about in terms of it's effects on the problem you want to teach the NN to solve?
Regards,
Thomas
Yes, that's definitely true and the root cause for this is that (artificial) neurons can only process scalar input data in terms of numbers.RuuKoo wrote:Do I really have to decode all to numerics before I can import raw CSV ?
What would you expect a date/time information to be converted during import or how should the NN process this information, respectively?
Additionally, date/time information most of the time is not a good type of information to be provided to an NN's inputs anyway, even if converted to a numeric value. This is due to the fact that date/time information by nature always evolves towards higher values and thus does not represent an information that can be used for generalization or for learning association rules: The date/time information by definition always differs from data set to data set.
The only way to incorporate date/time information in a possibly useful way is to convert it to some periodic value that represents a period which might be meaningful for the problem: E.g. a period of a year can be useful information for some problems that might be affected by time of year. In such a situation date/time information may be converted into a sine wave signal in periods of one year. This allows the NN to learn rules that are related to the time of year. Other problems may get along better with a time-of-day periodic signal.
What is your date/time information about in terms of it's effects on the problem you want to teach the NN to solve?
Regards,
Thomas
Thomas Jetter