imodels VS ANN-decompiler

- all domains requiring high interpretability absolutely ignore deep learning at all, and put all their research into traditional ML; see e.g. counterfactual examples, important interpretability methods in finance, or rule-based learning, important in medical or law applications

Next, this variable will be your target and you can use various supervised learning models to answer your question. Since interpretation is key, you can use something from here: https://github.com/csinva/imodels or do some black box models and use shab to understand which features contributed most.

Option 2) fit a model from https://github.com/csinva/imodels on the predicted values of the RF

Despite recent breakthroughs in the formulation and fitting of interpretable models, implementations are frequently challenging to locate, utilize, and compare. imodels solves this void by offering a single interface and implementation for a wide range of state-of-the-art interpretable modeling techniques, especially rule-based methods. imodels is basically a Python tool for predictive modeling that is simple, transparent, and accurate. It gives users a straightforward way to fit and use state-of-the-art interpretable models, all of which are compatible with scikit-learn (Pedregosa et al., 2011). These models can frequently replace black-box models while boosting interpretability and computing efficiency without compromising forecast accuracy. Continue Reading

Our package imodels is expanding our sklearn-compatible set of interpretable models and always looking for new contributors!

I can understand to why they would say that, as to what they refer as "woke" has massively invaded everything, but you are probably right that it was not done on purpose. Its not that easy to filter things like that out, especially not if humans have to do that manually. Also a friendly reminder that this is not real artificial intelligence, there is a good resource here https://github.com/Shamar/ANN-decompiler that explains this a bit better to why its mostly a magic trick, we have done models like that on paper in the 70's, not as large as those but its not exactly a new trick.

> Does this somehow smuggle the training dataset back into the VM?

Turns out you were right about this: http://www.tesio.it/2021/09/01/a_decompiler_for_artificial_n...

Obviously I was not aware of this, so the whole decompilation process was a waste of computation time, but it doesn't prove nor disprove anything about the "model"'s relation with the source dataset.

But I would be very happy to learn from you how to compute the whole source dataset from the output produced by compile.py without considering the "model".