How do multiple sensor types lead to more efficient neural encoding architectures?