Adapting to Structure and Using Structure to Adapt: Toward Explaining the Success of Modern Deep Learning