We need to decide the number of input neurons before designing the input layer. It can be derived from the feature shape. For instance, we have 13 input features (excluding the label). But after applying the transformation, we have a total of 11 feature columns present in the dataset. Noise features are removed and categorical variables are transformed during the schema transformation. So, the final transformed data will have 11 input features. There are no specific requirements for outgoing neurons from the input layer. If we assign the wrong number of incoming neurons at the input layer, we may end up with a runtime error:

The DL4J error stack is pretty much self-explanatory as to the possible reason. It points out the exact layer where it needs a fix (layer0, in the preceding example).