How to handle backward pass for split layer

[vamen]

Example in googlenet for each inception unit the input from previous layer is split into into input for three convolution unit.how to handle deconvolution in such cases ?