alibi
alibi copied to clipboard
SeldonIO
Algorithms for explaining machine learning models
The AnchorText with `sampling_strategy=language_model` uses Tensorflow language models from [transformers](https://github.com/huggingface/transformers). Some benchmarking showed that Tensorflow models are at least 1.5 slower than the corresponding Pytorch implementation. Thus, for full performance,...
AnchorText works with black-box models. Thus, the black-box can potentially be a transformer-based model. In this case, both transformers (the one corresponding to the black-box model and the one used...
The AnchorText language model extension supports the option of including a list of `stopwords`. This means that the words inside the list will not be perturbed. **Should we include the...
This would serve several purposes: - Realistic example using a common architecture - Performance benchmarking of the Anchor algorithm with realistic models - Uncover issues with GPU sharing #437 .
Currently `forward_kwargs` expects to contain arrays, this is to handle the use case of explaining `transformer` models. However, models can have more general `forward_kwargs` which should be handled. See [Captum...
Whilst `AnchorText` works directly on raw text, `IntegratedGradients` works on the token level. One reason for this is that `IntegratedGradients` is use case agnostic - tabular data, images and text...
The way `AnchorText` code finds word similarities is different between the two major spacy versions with implications on runtime and quality of words found. The scope of this would be...
This comes from the observation here: https://github.com/SeldonIO/alibi/pull/152#discussion_r428624088 Having done some performance tests, going via a `pandas` dataframe to do a groupby operation can take up to 25% of the computation...
Current ALE implementation only calculates first order effects, an extension to second order effects is possible. If `n` is the number of features then: - In ALE all `n` main...
ALE currently supports numerical features only. An extension to categorical features is possible, but comes with serious caveats for interpretability (see https://compstat-lmu.github.io/iml_methods_limitations/ale-misc.html) so I think some more research needs to...