A model, dataset agnostic method of mathematically analyzing the curvature of a neural network given it's weights and a test dataset, using Sheaf theory and Cech Cohomology - the computation variant of De Rhams Chomology which allows to find global maxims and minimums within a closed dataset.
The high dimensional sheafified output of curvature is tested by using b spline interpolation onto the input source which gives N^2 results of the batch these images are then merged using a merge Mertens algorithm and retrained on a model with it's weights been reset.
This is then tested with the original models as a control, and compared using a T-test.
Test results may be found in test_output.json.