Neural network Nonlinear regression

Hi,

I`d like to use Matlab deep learning tool for my research

problem is, there are few inputs and large outputs

my case, inputs are 4 numbers (e.g.) [0.1; -0.1; 0.1; 0.3]

and outputs are 99 numbers with range -10~10 (e.g) [0.2341; 0.56732; 1.52342; 6.4352;-1.4122; ...]

these inputs and outputs have some relationship extually (calculation is too complex, which is reason why i tried to use N.N)

please give some advice for

what network should i choose,

number of hidden layers,

number of nodes,

training function...etc

i want to make MSE of this stuff less then 1e-4

oh, i have tons of training data and targets already.

Answer 1

[Mahesh Taparia]

Hi

By looking at your input output size, I am assuming you are estimating 99 dimensional output from 4 dimensional input. You can use neural network, but not sure how much accurate it will be. Keep at least 5-6 hidden layer with leaky relu activations (as your prediction are negative also) and regression layer as output layer in order to reduce MSE. The number of nodes can be in increasing order from previous layer, for example you can have it in order->Input->8->16->32->64->99->output layer. Hope it will helps.

[tea hun kang]

Thanks for your kind advice, i`ll try it :)