# Avoid touching the Data Layer A powerful idea to keep in mind when dealing with Statistical Learning and AI systems is: > If you capture effective metadata representations of your underlying data you can build algorithms that allow you to avoid touching the underlying data until it is necessary. We are taking advantage of the intrinsic **structure** of the scenario, whether that may be the structure of zip codes and their relation to location on the earth (which we can use to relate to other locations on earth such as state, country, lat long, etc), or in the case of something like [SynFlow](SynFlow-Pruning-Neural-Networks-without-data.md) we are taking advantage of derived conservation laws of a network (a mathematical object). ### Examples * Let us have a dataset with the feature "zip code". This implicitly defines a *type*. If we capture this type, then given an [Ontology (Systems)](Ontology%20(Systems).md) we can see what types of transformations are available to us without ever needing to touch the data. For instance, a zip code could be converted into a lat-long pair. Or we could extract the corresponding state. * --- Date: 20211115 Links to: [Big Ideas MOC](Big%20Ideas%20MOC.md) Tags: References: * []()