Neural Network - Mapping of available inputs to corresponding outputs - Switching off unavailable input/output

I am just getting acquainted with neural networks toolbox of MATLAB. The feed forward network I need to build should be capable of switching off those input and output connections when there is no available data and correctly map the available data to the correct output.

For example,

My inputs are different salts/ions (Na, Mg, Cl, Ca, etc.) in a mixture and other parameters (temp, pressure, etc.) and my outputs are the performance of each ion. So say 10 input elements (6 ions + 4 other parameters). Then, the number of output elements will be 6 i.e. one corresponding to each ion. However, I have input data (vectors) that do not contain some of the ions. For example, one input pattern only has ions A and B (since only these two ions are in the solution mixture) but no ions C, D, E, or F. In this case, there should be no output from the nodes that correspond to C to F i.e. they should be switched off. Likewise, the second input pattern may have ions C and F, and the third may have all 6 ions.

How can I create a network that will recognize which input ions correspond to which output ions and effectively switch off the input element (that have unavailable data) and corresponding output element? I have to present the same size input vector every time (10 elements). Switching off input and output neurons means their weights to the hidden layer will not be used in the forward or backward passes and these weights will retain have to their previous values.

Thank you.

Answer 1

[Greg Heath]

Train the net with sample input/output pairs. Use zeros to indicate the no input and no output states.

Greg