The idea is simple.

If a problem of high complexity requires a complex weight matrix. Then I think you can just iterate and ?determine it by using Gaussian filter in the process of creating the weight matrices.

Then by looking at the softened weight matrices or using them in a classification network you then get a number or label of the ?complexity. With this I guess you can then later adapt a non smoothed network with the right model. That is. This method determines if the problem is easy or hard.

If you classify the different complexity and its corresponding model setup I believe a method similar to this one can help solve hard problems

Update

**If anything would solve the complexity problem it would be a network.**

So I thought if I get a smooth converged weight matrix. Then train on an additional set of samples from the test set and then compare the change of the weight matrix to the original. I guess this would reflect how difficult the problem is. The ?more change. The more problematic the problem was.

Could this be a way to predict how hard a problem is. I mean. In the brain we think a little bit about the problem then decide if its difficult or not. So inspired by this. Could a machine learning network solve this question?