Improving Compositionality of Neural Networks by Decoding Representations to Inputs