sstanom

Benchmark: SST Anomaly

This benchmark is an implementation of Mamalakis et al. (2022), with additional features.

The purpose is to quantitatively compare XAI methods based on the correlation between the XAI output and the attribution of a known function F. The synthetic function F is carefully designed to have (1) spatial relationships among the grid cells and (2) the ability to calculate the attribution of each grid cell toward the function's output. By generating a very large amount of synthetic samples, a neural network is trained to approiximate F. Since the NN achieves near-perfect performance (R-square > 0.999), it is assumed that the learned function is a good approximation of the known function F. So, differences between the attribution of F and the output of XAI methods is assumed to be because of limitations in the XAI method rather than differences between what the model learned and F.

Benchmark Pipeline Diagram

Image from Mamalakis et al. (2022)

Benchmark Choices

This code is designed to create synethetic benchmarks from four base data sets. Each created by extracting samples from the COBE-SST 2 and Sea Ice Gridded Climate Dataset. This benchmark folder is called sstanom since that was the original purpose.

1. sst : Sea Surface Temperature, directly from COBE.
2. icec : Sea Ice Concentration, directly from COBE.
3. sstanom : SST Anomaly, created from SST by subtracting the climatological mean and detrending.
4. icecanom : ICEC Anomaly, created from ICEC by subtracting the climatological mean and detrending.

How to Run

To create a synthetic benchmark and run XAI, you need to setup a configuration file and run the pipeline command.

Configuration file

Below is a configuration file for the sstanom pipeline that replicates the results of Mamalakis et al. (2022)

{
    "out_dir": "benchmarks/sstanom/out/sstanom/",
    "covariance_file": "benchmarks/sstanom/out/sstanom/cov.npz",
    "n_samples": "1000000",
    "n_pwl_breaks": "5",
    "samples_to_plot": "0;10;100;200;300",
    "pwl_functions_to_plot": "0;10;100;200",
    "nn_hidden_nodes": "512;256;128;64;32;16",
    "nn_epochs": "50",
    "nn_batch_size": "32",
    "nn_learning_rate": "0.02",
    "nn_validation_fraction": "0.1",
    "xai_methods": "saliency;integrated_gradients;input_x_gradient;lrp",
    "variable": "sst",
    "anomaly": "true"
}

The following describes each option. Read the Mamalakis et al. (2022) paper to get more insight on their purpose. These short descriptions are simply to help you match them up to the concepts in the paper.

• out_dir : Path to save all pipeline output files.
• covariance_file : Where the covariance file is. It gets created by this pipeline.
• n_samples : Number of synthetic samples to generate using the covariance matrix.
• n_pwl_breaks : Number of breakpoints in the piece-wise linear functions that make the known function F.
• samples_to_plot : While you can plot samples later, its useful to plot a few during the pipeline for debugging.
• pwl_functions_to_plot : It is also useful to check a few of the piece-wise linear functions.
• nn_hidden_nodes : Specifiy the neural net's hidden layers.
• nn_epochs : Number of training epochs.
• nn_batch_size : Batch size for neural network training.
• nn_learning_rate : Learning rate for neural network training.
• xai_validation_fraction : Fraction of synthetic samples to use as validation.
• xai_methods : Which XAI methods to run.
• variable : Which dataset to download and use from COBE data source. Choices are (sst, icec)
• anomaly : Whether or not to convert to anomaly data (subtract climatology and detrend)

Available XAI methods

See the Mamalakis et al. (2022) paper for method descriptions.

• saliency : saliency maps
• integrated_gradients : integrated gradients
• input_x_gradient : Input X gradient
• lrp : LRP-zero

Pipeline

# Example: sst anomaly pipeline
bash benchmarks/sstanom/create_sstanom_benchmark.sh benchmarks/sstanom/config_sstanom.json

There is also a script to plot several attribution maps to compare

dir=benchmarks/sstanom/out/sstanom/

python src/plot/plot_attributions.py \
    --attr_files  $dir/pwl-out.npz,$dir/xai/xai_input_x_gradient.npz,$dir/xai/xai_integrated_gradients.npz,$dir/xai/xai_saliency.npz \
    --sample_idxs 0,10,100,200,300.0,1,2,3,4.0,1,2,3,4.0,1,2,3,4 \
    --names ground_truth,input_x_grad,integrated_grad,saliency