Information Maximization Perspective of Orthogonal Matching Pursuit with Applications to Explainable AI
Aditya Chattopadhyay,¹ Ryan Pilgrim,¹ and René Vidal²
¹Johns Hopkins University, USA, {achatto1,rpilgri1} <at> jhu.edu
²University of Pennsylvania, USA, vidalr <at> seas.upenn.edu
This repository contains code to accompany Information Maximization Perspective of Orthogonal Matching Pursuit with Applications to Explainable AI (NeurIPS 2023).
Information Pursuit (IP) is a classical active testing algorithm for predicting an output by sequentially and greedily querying the input in order of information gain. However, IP is computationally intensive since it involves estimating mutual information in high-dimensional spaces. This paper explores Orthogonal Matching Pursuit (OMP) as an alternative to IP for greedily selecting the queries. OMP is a classical signal processing algorithm for sequentially encoding a signal in terms of dictionary atoms chosen in order of correlation gain. In each iteration, OMP selects the atom that is most correlated with the signal residual (the signal minus its reconstruction thus far). Our first contribution is to establish a fundamental connection between IP and OMP, where we prove that IP with random projections of dictionary atoms as queries "almost" reduces to OMP, with the difference being that IP selects atoms in order of normalized correlation gain. We call this version IP-OMP and present simulations indicating that this difference does not have any appreciable effect on the sparse code recovery rate of IP-OMP compared to that of OMP for random Gaussian dictionaries. Inspired by this connection, our second contribution is to explore the utility of IP-OMP for generating explainable predictions, an area in which IP has recently gained traction. More specifically, we propose a simple explainable AI algorithm which encodes an image as a sparse combination of semantically meaningful dictionary atoms that are defined as text embeddings of interpretable concepts. The final prediction is made using the weights of this sparse combination, which serve as an explanation. Empirically, our proposed algorithm is not only competitive with existing explainability methods but also computationally less expensive.
This project uses the conda package manager, available here.
If on Mac, type
conda env create -f environment_mac.yml
If on Linux, type
conda env create -f environment_linux.yml
Once this command has finished, you can activate the environment with conda activate.
Windows is currently not supported.
To execute all of the sparse recovery experiments in one go, run
./script/sparse_recovery_experiments.sh
from the repository root directory. To plot all of the results from running the above script, run
./script/sparse_recovery_plots.sh
Due to the way things are saved, these scripts will consume a lot of disk space (≈900 GB). We are working on lightening this disk usage.
To run a smaller subset of experiments, you can use the ip_omp/simulations.py CLI:
$ python -m ip_omp.simulations --help
Usage: python -m ip_omp.simulations [OPTIONS] RESULTS_DIR
╭─ Arguments ──────────────────────────────────────────────────────────────────╮
│ * results_dir PATH Where to save the results. [default: None] │
│ [required] │
╰──────────────────────────────────────────────────────────────────────────────╯
╭─ Options ────────────────────────────────────────────────────────────────────╮
│ --problem-size [small|large] The size of the │
│ problem n=256 │
│ (small) or n=1024 │
│ (large). │
│ [default: small] │
│ --coeff-distribut… [sparse_gaussian| The distribution │
│ sparse_const] of the │
│ coefficients of │
│ the sparse code. │
│ [default: │
│ sparse_gaussian] │
│ --noise-setting [noiseless|noisy] The noise setting │
│ of the │
│ experiment. │
│ [default: │
│ noiseless] │
│ --overwrite --no-overwrite Whether to │
│ overwrite │
│ existing results. │
│ [default: │
│ no-overwrite] │
│ --jobs INTEGER RANGE Maximum number of │
│ [x>=1] subprocesses to │
│ run. │
│ [default: 1] │
│ --device [cuda|cpu] Device to use │
│ (CPU/GPU). │
│ [default: cpu] │
│ --memory-usage FLOAT RANGE Memory usage │
│ [0.0<=x<=1.0] below which a GPU │
│ will be │
│ considered free │
│ when launching │
│ new subprocesses. │
│ [default: 0.75] │
│ --utilization FLOAT RANGE Utilization below │
│ [0.0<=x<=1.0] which a GPU will │
│ be considered │
│ free when │
│ launching new │
│ subprocesses. │
│ [default: 0.75] │
│ --order-by [utilization|memo How to sort │
│ ry_usage] available GPUs │
│ when launching │
│ subprocesses. │
│ [default: │
│ utilization] │
│ --help Show this message │
│ and exit. │
╰──────────────────────────────────────────────────────────────────────────────╯
For finer-grained control over plotting, you can also use the ip_omp/figures.py CLI. The plot-noiseless command is meant for plotting the noiseless experiment results:
$ python -m ip_omp.figures plot-noiseless --help
Usage: python -m ip_omp.figures plot-noiseless [OPTIONS] SMALL_RESULT_PATH
LARGE_RESULT_PATH
Plot the results of the noiseless sparse recovery experiments.
╭─ Arguments ──────────────────────────────────────────────────────────────────╮
│ * small_result_path PATH Path to the small (n=256) results file. │
│ [default: None] │
│ [required] │
│ * large_result_path PATH Path to the large (n=1024) results file. │
│ [default: None] │
│ [required] │
╰──────────────────────────────────────────────────────────────────────────────╯
╭─ Options ────────────────────────────────────────────────────────────────────╮
│ --coeff-distribut… [sparse_gaussian| The distribution │
│ sparse_const] of the │
│ coefficients of │
│ the sparse code. │
│ [default: │
│ sparse_gaussian] │
│ --save-dir PATH Where to save │
│ resulting plots. │
│ [default: .] │
│ --save-file-format [eps|png] Format of the plot │
│ file. │
│ [default: png] │
│ --together --no-together Whether to plot │
│ small (n=256) and │
│ large (n=1024) │
│ results in the │
│ same figure. │
│ [default: │
│ no-together] │
│ --help Show this message │
│ and exit. │
╰──────────────────────────────────────────────────────────────────────────────╯
The plot-noisy command is meant for plotting the noisy experiment results:
$ python -m ip_omp.figures plot-noisy --help
Usage: python -m ip_omp.figures plot-noisy [OPTIONS] SMALL_RESULT_PATH
LARGE_RESULT_PATH
Plot the results of the noisy sparse recovery experiments.
╭─ Arguments ──────────────────────────────────────────────────────────────────╮
│ * small_result_path PATH Path to the small (n=256) results file. │
│ [default: None] │
│ [required] │
│ * large_result_path PATH Path to the large (n=1024) results file. │
│ [default: None] │
│ [required] │
╰──────────────────────────────────────────────────────────────────────────────╯
╭─ Options ────────────────────────────────────────────────────────────────────╮
│ --save-dir PATH Where to save resulting plots. │
│ [default: .] │
│ --save-file-format [eps|png] Format of the plot file. [default: png] │
│ --help Show this message and exit. │
╰──────────────────────────────────────────────────────────────────────────────╯
The code supports 5 image classification datasets: Imagenet, Places365, Cifar10, Cifar100, CUB200.
All datasets are to be downloaded in the ./ip_omp/data/{dataset} directory, where {dataset} is replaced by the name of the dataset the user is interested in.
Cifar10, Cifar100 and Places365 are avaiable as part of the torchvision datasets and the provided codes would automatically download this for the user.
Imagenet - refer to instructions at https://www.image-net.org/download.php. Remember to save the dataset at "./data/ImageNet".
CUB200 - refer to instructions at https://www.vision.caltech.edu/datasets/cub_200_2011/. Remember to save the dataset at "./ip_omp/data/CUB/", with the image directory being located at "./ip_omp/data/CUB/CUB_200_2011".
The code for CLIP-IP-OMP algorithm is provided in the file clip_ip_omp.py. The code process the entire training and test set for the input dataset and saves the IP-OMP code for each image into a .npy file in the saved_files directory.
Usage: python -m ip_omp.clip_ip_omp -s {sparsity level} -bs {batch size} -dataset {dataset name}
- sparisty level (integer) is the desired number of non-zero coefficients in the sparse code. This controls the complexity of the explanations.
- batch-size (integer) is the number of images to process in one batch. Default value is 128.
- dataset name is a value from the set {"imagenet", "places365", "cub", "cifar10", "cifar100"}
Due to the large size of the Imagenet and Places365 dataset. We first preprocess every image in the dataset to obtain their corresponding CLIP image features. This is done via the preprocess.py code. Run this code before running clip_ip_omp.py for these dataset.
Usage: python -m ip_omp.preprocess -dataset {dataset name}
- dataset name is a value from the set {"imagenet", "places365"}
This is done using the train_linear_classifier.py code. Run this after running the clip_ip_omp.py code.
Usage: python -m ip_omp.train_linear_classifier -s {sparsity level} -bs {batch size} -dataset {dataset name}
- sparisty level (integer) is the sparsity level used while constructing the IP-OMP codes using the clip_ip_omp.py file.
- batch-size (integer) is the number of images to process in one batch. Default value is 128.
- dataset name is a value from the set {"imagenet", "places365", "cub", "cifar10", "cifar100"}
This repository is MIT-licensed. See LICENSE for details.
@inproceedings{
chattopadhyay2023information,
title={Information Maximization Perspective of Orthogonal Matching Pursuit with Applications to Explainable {AI}},
author={Aditya Chattopadhyay and Ryan Pilgrim and Ren\'{e} Vidal},
booktitle={Thirty-seventh Conference on Neural Information Processing Systems},
year={2023},
url={https://openreview.net/forum?id=CAF4CnUblx}
}