The typical pipeline in ML is
- Find a data-related problem that you want to solve
- Build a model or algorithm that feeds on data related to the problem to try to solve the problem
- Check if the solution produced by the model or algorithm is satisfactory by testing the prediction against another piece of data
My question is: given the vast amount of models, algorithms, and data currently available to use, e.g. on repositories like github etc, are there any current efforts that follow an inverted pipeline as follows?
- Choose an existing model or algorithm or create a new one
- Test it against many different types of data sets (as long as input output formats are compatible
- Find problems for which the model or algorithm is good at, even if it was originally not design for this domain, or designed with any purpose at all
In other words, systematically test existing algorithms, or a new algorithm, against existing datasets to find triples of (algorithm, question, data) that work well and have not been thought together before.
I would image this to be increasingly easy to automate using things like autoML and etc.
Just to clarify: I'm not so much looking for a list of reasons why this is a bad idea, or why it can, or cannot be done (although I will take advice if given and good). I just want some references of people or papers doing something similar to this, if any. A reference that partially follows the inverted pipeline is better than nothing.
Thank in advance