How accurate does a model-based prior over search directions have to be?
We rotated the analytic Jacobian prior to test this. At 90°, it fails as expected.
But surprisingly, up to 30° the learning speed is barely affected, and final performance stays unchanged even up to 45°.
