With the emergence of Big Data, handling incomplete data sets has also become an interest area. This is especially true in the areas of human provided data. One way to deal with such data sets that includes incomplete input data is to create a smaller, complete subset of the input data, but this approach comes with its own disadvantages. The aim of thesis is to provide a novel algorithm for training neural networks with incomplete input data sets. The proposed method uses masking of weights in order to partially train the network, depending on the available input. In effect, this enables updating only the weights related to the available input. The experimental results show that this method could be viable approach when dealing with less than 3-5% missing data.