A code-generating model to output accurate results for complex visual input tasks