Hi, I am not sure the underlying technology of ludwig to generate a trained model, e.g., image classification. Is it similar to some kind of netural architecture search? Are there papers or some docs to give more descriptions?
Hi AllenDun, thanks for your question.
So at the moment there's nothing like that, Ludwig just has some decent default parameters for all the encoders that make it work decently in most of the cases, then to make it work really well, you have to optimize your parameters manually.
At the same time at Uber AI there are a bunch of great researchers working on neuroevolution, bayesian optimization and discrete optimization. Internally we have some tooling to optimize the hyperparameters of the models in a parallel distributed way, but they are tied to our infrastructure, so they would be worthless if released externally. But we are investigating ways to do a mix of hyperparameter and architecture search that is disconnected from our infrastructure. We are still in early stages on this, but stay tuned, it may eventually be released as open source either as part of Ludwig or as a separate package that optimizes Ludwig models.