I am creating an ANN which has 3 input neurons which take inputs from the device "s accelerometer in the form of x , y , z. These values are positive as well as negative depending upon the acceleration. I am not able to get an suitable activation to normalize these values. Also , I am not getting desired predictions. Any help will be valuable. :-)
Asked
Active
Viewed 161 times
1 Answers
0
You should normalize your data before. I would advise using standard score, which looks as following:
- collect training sets for each of the input variables
- calculate mean (m) and standard deviation (std) for each of the sets
- normalize as (x-m)/z
If you are working on a regression problem, don't forget to normalize target values as well.
you can also use other normalization techniques if you think they would work better for your case. Some of them you can see here.
Choice of the activation function, in this case, should not affect much, you can just play with different types and see which results in a better performance.
asakryukin
- 2,524
- 1
- 13
- 14
-
Thanks for your help. :-) – Jun 08 '18 at 05:28
-
Also , can I compress the input values between -1 and 1 using Hyperbolic tangent function ? Will that be helpful ? – Jun 08 '18 at 05:31
-
@user9477964 I wouldn't say it is a good idea, first of all, tanh is non-linear, so you can lose some important relationships between input values, as -100 and -10 will basically become 1 and that's not what you want. Activation function should be used as an instrument for introduction of non-linear properties to the network, but not as a normalization function. – asakryukin Jun 08 '18 at 08:45