Invariance of retinal output during visual learning