Neural network training function error modulo a number

How can I tell neural network training function to compute the error modulo a number.? For example if my targets are angles in degrees and the desired is 359 but the predictor gave me 2 degrees, the error is large (359-2)^2. How can you tell neural network to treat 2 as 360-2, such that the error is now (359-358)^2?

Answer 1

[Greg Heath]

The simplest function that makes perfect sense is the corresponding squared Euclidean distance between corresponding points on the unit circle.

Other than taking the square root to obtain linear distance, I see no reason to waste computation to use any other function of distance,

Hope this helps.

Thank you for formally accepting my answer

Greg

[Kostas Sarrigeorgidis]

Are u saying instead of train on the angle (one output), train with cos(theta), sin(theta)? (2 outputs). If so, how do you tell the network to give answers on the unit circle?