This repository contains the code for the experiments included in the paper Kernelized Adversarial Concept Erasure, accepted in EMNLP 2022.
In the paper, we run a minimax concept erasure game in kernel space, in which we aim to prevent a given kernel classifier from reconstructing a given concept (e.g. gender) from neural representations.
Our method is based on applying Nystrom feature mapping to approximate the kernel features, following by the adversarial formulation presented in the papr Linear Adversarial Concept Erasure (RLACE).
The file run_kernels.py' contains the function calc_nystrom' to calculate the nystrom features for a given kernel, and the function calc_preimage_nystrom_mse' that calculates an approximate preiamge of the clean features in kernel space (after the removal of, e.g., gender-encoding features). The file relaxed_inlp.py' contains an implementation of the RLACE algorithm that we run over the nystrom features.
We rely on the package MKLpy for multiple kernel learning.
First, run the file `run_kernels.py' in order to calcualte the nystrom mapping, run the RLACE algorithm, and store the approximate preimage features:
python3 run_kernels.py --run_id 1 --mode glove --device 0 --normalize 1
Then, run the script `eval_kernels.py' in order to train new kernel adversaries over the preimage features, and record their performance.
python3 eval_kernels.py --run_id 1 --normalize 1
The notebook eval_kernels.ipynb' collects the accuracies of the different kernel adversaries and generates the accurcy tabels in the paper. The notebook additional_analysis.ipynb' contains a semantic analysis of the clean preimage vectors -- specifically, WEAT, closet-neighbors test and visualization.