acd 0.0.1

Hierarchical interpretatations and contextual decomposition in pytorch

Official code for using / reproducing ACD from the paper Hierarchical interpretations for neural network predictions (ICLR 2019 pdf). This code produces hierarchical interpretations for a single prediction made by a neural network.

Note: this repo is actively maintained. For any questions please file an issue.

examples/documentation

• installation: pip install git+https://github.com/csinva/hierarchical-dnn-interpretations
• examples: the reproduce_figs folder has notebooks with many demos
• api: the api gives a list of available functions
• src: the acd folder contains the source for the method implementation
• allows for different types of interpretations by changing hyperparameters (explained in examples)
• tested with python3 and pytorch >1.0 with/without gpu
• all required data/models/code for reproducing are included in the dsets folder

Inspecting NLP sentiment models	Detecting adversarial examples	Analyzing imagenet models

notes on using ACD on your own data

• the current CD implementation doesn't always work for all types of networks. If you are getting an error inside of cd.py, you may need to write a custom function that iterates through the layers of your network (for examples see cd.py). Should work out-of-the-box for many common layers though, including antyhing in alexnet, vgg, or resnet.
• to use baselines such build-up and occlusion, replace the pred_ims function by a function, which gets predictions from your model given a batch of examples.

related work

• PDR framework (PNAS 2019 pdf) - an overarching framewwork for guiding and framing interpretable machine learning
• CDEP (ICML 2020 pdf, github) - penalizes CD / ACD scores during training to make models generalize better
• TRIM (ICLR 2020 workshop pdf, github) - using simple reparameterizations, allows for calculating disentangled importances to transformations of the input (e.g. assigning importances to different frequencies)
• DAC (arXiv 2019 pdf, github) - finds disentangled interpretations for random forests
• Baseline interpretability methods - the file scores/score_funcs.py also contains simple pytorch implementations of integrated gradients and the simple interpration technique gradient * input

reference

• feel free to use/share this code openly
• if you find this code useful for your research, please cite the following:

@inproceedings{
   singh2018hierarchical,
   title={Hierarchical interpretations for neural network predictions},
   author={Chandan Singh and W. James Murdoch and Bin Yu},
   booktitle={International Conference on Learning Representations},
   year={2019},
   url={https://openreview.net/forum?id=SkEqro0ctQ},
}