The compositionality of neural networks: integrating symbolism and connectionism