R neural net package hidden neurons, single vs vector argument

user_6093015 0
Viewed 524

In R neural net package, what is the difference between a single number argument and a vector argument for hidden neurons? e.g. hidden = 100 vs hidden = c(100,50)

data-science deep-learning machine-learning neural-network r
1 Answers

You use it to provide 2 information, the number of layers, which is the length of the vector, and the number of neurons per layer.

So in your question, neuralnet(..hidden=100) means 1 layer with 100 neurons, whereas neuralnet(..hidden=c(100,50)) means 2 layers, 1st layer 100 and 2nd layer 50.

We can see this by looking at the results matrix which contains the estimated weights of the connections:

fit = neuralnet(Species == "setosa" ~ ., iris, linear.output = FALSE,hidden=2)
fit$result.matrix

                                         [,1]
error                             0.008141255
reached.threshold                 0.008900248
steps                            48.000000000
Intercept.to.1layhid1            -0.996032931
Sepal.Length.to.1layhid1          0.433339769
Sepal.Width.to.1layhid1          -1.178668127
Petal.Length.to.1layhid1         -0.483878954
Petal.Width.to.1layhid1           4.399508265
Intercept.to.1layhid2             1.764932246
Sepal.Length.to.1layhid2         -0.748650261
Sepal.Width.to.1layhid2           3.602604148
Petal.Length.to.1layhid2         -2.226669079
Petal.Width.to.1layhid2          -3.210541627
Intercept.to.Species == "setosa" -1.422643815
1layhid1.to.Species == "setosa"  -3.596971312
1layhid2.to.Species == "setosa"   6.205140270

So you have only 1 layer and 2 hidden neurons in that layer which are connected to the variables and output layer.

If you do:

fit = neuralnet(Species == "setosa" ~ ., iris, linear.output = FALSE,hidden=c(2,1))
fit$result.matrix

                                         [,1]
error                             0.010978699
reached.threshold                 0.007306809
steps                            92.000000000
Intercept.to.1layhid1             4.616674345
Sepal.Length.to.1layhid1          3.342501151
Sepal.Width.to.1layhid1           4.367109042
Petal.Length.to.1layhid1          5.385330435
Petal.Width.to.1layhid1           1.402765434
Intercept.to.1layhid2            -3.828104421
Sepal.Length.to.1layhid2          1.013023746
Sepal.Width.to.1layhid2          -6.037198731
Petal.Length.to.1layhid2          3.705741605
Petal.Width.to.1layhid2           7.993464809
Intercept.to.2layhid1            -2.075821677
1layhid1.to.2layhid1             -2.550031126
1layhid2.to.2layhid1              9.393164468
Intercept.to.Species == "setosa"  4.146671533
2layhid1.to.Species == "setosa"  -8.957645357

It's similar to before except now you have a 2nd layer with one neurons and this is connected to the output layer.

user_122584590
Related