I’ve just started coding my NN after a month or two of reading.
after tweaking some models and spending time on some trials&errors with keras,
I think I’ll to stick to this NN (described below), but to my knowledge, so far, it’s neither a classification, nor a regression.
the number of inputs (samples) : 20,000
number of features per input : 1500
number of outputs : 7 (but all are float values in range [-1,1])
e.g: a possible output is : (0.2, 0.5, 0.0, -0.6, 0.8, -0.3, 1)
is this 7 outputs, or 1 output ?
I’m having a hard time, trying to understand how my loss, and accuracy is gonna be computed/used.
output is not a single float (a scalar value),
not a 2D point (a dot on x,y plane),
not a 3D point (a point on x,y,z space)
but it’s a seven-float thing. (a point in 7D space? )
can I just use this 7D points as targets, and train on them using usual MSE and MAE and other similar loss methods ?
I know I can’t use Softmax activation at last layer, and since I can’t define targets as classes, class weights can’t be calculated. So it definitely is not a classification problem.
but, how do I build a valid model for such output ?
(based on some readings, is it correct to not to use a activtion function on last layer ?)