Simple low-dimensional computations explain variability in neuronal activity