Machine Learning Idea – Do Cells Solve Problems By Splitting?

I think one problem with deep learning is that if you add an ?odd looking training sample to the model(). Then it could ?perhaps deflect a little. Give off a little on the prediction accuracy.

So the idea is to check what the response new samples have on the already learned samples. Maybe its possible to have a system that senses deflection in the weights. One idea would be split the problem into two units and so on? Something similar to cells?