1
$\begingroup$

Is it possible to design a network with an unknown number of neurons in the output layer?

I am trying to solve a classification problem, where I use motorcycles' exterior color, interior color, and make to predict the type of damages that are going to show on them.

Since I don't know what the possible damage combinations are going to look like (ex: 2 scratches and 1 scuff or 3 scratches and 1 scuff), then it is not possible for me to know how many output neurons I have in my network.

This is my first ever network. I am trying to imagine what the input and output layers look like. The output layer seems to contain "n" neurons.

Is there a solution for this?

$\endgroup$
1
$\begingroup$

To deal with this as a classification problem you need to define your possible class combinations before hand, else how would you define your training ground truth.

It seems like you are dealing with a regression problem here since you are estimating the amount of two variables scratch and scuff, rather than if a scratch of scruff exist or not, and should try formulating it as such.

$\endgroup$
  • $\begingroup$ I am not trying to estimate the amount. I am determining what the actual damages are by feeding millions of records and mapping from the motorcycles' exterior color, interior color, and make to the combination of damages. My problem is that there is a number of neurons that I have to put in the output layer, which are going to be many. The reason obviously is because you can have many many combinations like: 3 scratches, or 2 scratches, or 2 scratches and 1 ding. There are 100 types of damages. So,I could have thousands of different combinations. $\endgroup$ – M J May 18 '16 at 14:31
0
$\begingroup$

If there are 100 types of damages, then each combination can be represented as "$n_1$ scratches, $n_2$ dings, $n_3$ damages of type 3 ... $n_{100}$ damages of type 100". Your task is to estimate $n_1$, ..., $n_{100}$ from the input data. This leads to a neural network which has 100 output neurons. Unless only one damage of type $t$ can exist, the activation function for the unit $t$ should be linear rather than log-sigmoidal.

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.