For reproducibility of results it would be good to store a unique identifier of the neural network models used to produce them. Many groups may be training models, and we need a way to match samples to trained models.
One idea would be to calculate a hash of a model (or its weights?) which could be saved in the metadata of a samples file (and the model file?). It would be good to have a way to go from this identifier to the model itself, though, so maybe we need to have a table of models and hashes. But then who maintains this table? How long does it take to calculate such a hash?
I am open to suggestions on the best way to solve these problems. Any thoughts?
For reproducibility of results it would be good to store a unique identifier of the neural network models used to produce them. Many groups may be training models, and we need a way to match samples to trained models.
One idea would be to calculate a hash of a model (or its weights?) which could be saved in the metadata of a samples file (and the model file?). It would be good to have a way to go from this identifier to the model itself, though, so maybe we need to have a table of models and hashes. But then who maintains this table? How long does it take to calculate such a hash?
I am open to suggestions on the best way to solve these problems. Any thoughts?