Divisive Normalization, Line Attractor Networks and Ideal Observers