shap VS captum

something like GitHub - shap/shap: A game theoretic approach to explain the output of any machine learning model.?

A less than 100 lines of code implementation of KernelSHAP because I had a hard time understanding shap's code.

We use tools like SHAP to explain what the model is doing to stakeholders.

https://github.com/slundberg/shap this or https://github.com/marcotcr/lime would be relevant to you, especially if you want to look at explaining a single prediction.

See this closed topic for more detail: https://github.com/slundberg/shap/issues/29

Dr. Molnar recently had a semi-viral post on LinkedIn and on Twitter, where he essentially highlights the booming popularity [and power] of using SHAP for explainable AI (which I agree with), but that it also comes with problems; i.e., the open source implementation has thousands of pull requests, bugs, and issues and yet there is no permanent or significant funding to go in and fix them.

Option 4) create SHAP values https://github.com/slundberg/shap to better understand what the RF did.

txtai pipelines are wrappers around Hugging Face pipelines with logic to easily integrate with txtai's workflow framework. Given that, we can use the SHAP library to explain predictions.

Most likely, NNs in general love shortcut learning (see Geirhos et al. 2020). In general, local explanations such as grad-cam are quite noisy, and sometimes even inconsistent (see Seo et al. 2018 ). Now, in my experience, I've seen that integrated gradients (see Sundararajan et al 2017) does a better job compared to Grad-CAM (also, add a noise tunnel), but this is only based on my limited experience. I would totally recommend using the implementations from the Captum library for loca explanations.

Take a look at captum

Sometimes. If explainable predictions are part of your business requirements, it's probably better not to rely entirely on black box models and instead design a system that gives you the information you need as part of it. If you end up using black box models, there are still methods that attempt to help attribute explanations to your prediction. Here's an example of a toolkit for attributing explanations post-hoc to black box model predictions: https://github.com/pytorch/captum

The straightforward way to interpret RL agent's decision is to use captum library.

My use-case is research-oriented. I work on Explainable AI. Generally, the best package I've come across to compute attributions in Pytorch Captum. If your object detector is in PyTorch, you can perhaps build it in.

Do any JAX experts know if there is an equivalent to https://captum.ai/ - a model interpretability library for pytorch?

In particular i want to be able to measure feature importance on both inputs and internal layers on a sample by sample basis. This is the only thing currently holding me back from using JAX right now.

Alternatively a simle to read/understand/port implementation of DeepLIFT would work too.

thanks

I'm interested to use JAX but am having a hard time finding anything similar to captum for the pytorch world.

Do you actually want to implement something? There are decent explainability libraries now, e.g., [AIX360](https://aix360.mybluemix.net/), [InterpretML](https://interpret.ml/), or [captum](https://captum.ai/). Pytorch + maybe pytorch lightning + captum might be the quickest way to actually implement something like an explainable neural net yourself. Do the standard tutorials for each of them and watch a few YT videos (or follow a coursera course or something like that) about how these things work in theory and practice, and you'll get up to speed relatively quickly. You will not be able to do useful work in ML without actually learning the ropes.