0
$\begingroup$

you have been hired by a gem mining company to develop a classification system that can classify gems as part of the automated sorting system.

you decided to use a network with one hidden layer. how would you go about determining the best number of hidden units to use in this layer

$\endgroup$
  • $\begingroup$ What does the input data look like? Is it an image or a set of quality variables? Why did you decide to use just one hidden layer? $\endgroup$ – Thomas Lumley Jun 11 at 7:56
  • $\begingroup$ in input data is color - Represented by RGB component between 0.0 and 1.0 for each color. brightness - between 0.0 and 1.0. flaws 0 to 10 worst. the quality grade ranges from A(best) b, c and D(worst) $\endgroup$ – Nivan Desilva Jun 11 at 8:07
0
$\begingroup$

Since that is very little information to go on, assuming the 1 hidden layer is fixed, i would start with n_hidden = max(n_in, n_out) neurons in the layer and then both increase and then decrease the neurons in steps all while monitoring the relevant metric using cross validation. The basic formula comes from Introduction to Neural Networks with Java by Jeff Heaton and has served me well so far.

| cite | improve this answer | |
$\endgroup$

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