An implementation of Relaxed Linear Adversarial Concept Erasure (RLACE)

Related tags

Machine LearningRLACE
Overview

Background

This repository contains an implementation of Relaxed Linear Adversarial Concept Erasure (RLACE). Given a dataset X of dense representations and labels y for some concept (e.g. gender), the method identifies a rank-k subsapce whose neutralization (suing an othogonal projection matrix) prevents linear classifiers from recovering the concept from the representations.

The method relies on a relaxed and constrained version of a minimax game between a predictor that aims to predict y and a projection matrix P that is optimized to prevent the prediction.

How to run

A simple running example is provided within rlace.py.

Parameters

The main method, solve_adv_game, receives several arguments, among them:

  • rank: the rank of the neutralized subspace. rank=1 is emperically enough to prevent linear prediction in binary classification problem.

  • epsilon: stopping criterion for the adversarial game. Stops if abs(acc - majority_acc) < epsilon.

  • optimizer_class: torch.optim optimizer

  • optimizer_params_predictor / optimizer_params_P: parameters for the optimziers of the predictor and the projection matrix, respectively.

Running example:

num_iters = 50000
rank=1
optimizer_class = torch.optim.SGD
optimizer_params_P = {"lr": 0.003, "weight_decay": 1e-4}
optimizer_params_predictor = {"lr": 0.003,"weight_decay": 1e-4}
epsilon = 0.001 # stop 0.1% from majority acc
batch_size = 256

output = solve_adv_game(X_train, y_train, X_dev, y_dev, rank=rank, device="cpu", out_iters=num_iters, optimizer_class=optimizer_class, optimizer_params_P =optimizer_params_P, optimizer_params_predictor=optimizer_params_predictor, epsilon=epsilon,batch_size=batch_size)

Optimization: Even though we run a concave-convex minimax game, which is generallly "well-behaved", optimziation with alternate SGD is still not completely straightforward, and may require some tuning of the optimizers. Accuracy is also not expected to monotonously decrease in optimization; we return the projection matrix which performed best along the entire game. In all experiments on binary classification problems, we identified a projection matrix that neutralizes a rank-1 subspace and decreases classification accuracy to near-random (50%).

Using the projection:

output that is returned from solve_adv_game is a dictionary, that contains the following keys:

  1. score: final accuracy of the predictor on the projected data.

  2. P_before_svd: the final approximate projection matrix, before SVD that guarantees it's a proper orthogonal projection matrix.

  3. P: a proper orthogonal matrix that neutralizes a rank-k subspace.

The ``clean" vectors are given by X.dot(output["P"]).

Owner
Shauli Ravfogel
Graduate student, BIU NLP lab
Shauli Ravfogel
A collection of video resources for machine learning

Machine Learning Videos This is a collection of recorded talks at machine learning conferences, workshops, seminars, summer schools, and miscellaneous

Dustin Tran 1.5k Dec 29, 2022
An open source framework that provides a simple, universal API for building distributed applications. Ray is packaged with RLlib, a scalable reinforcement learning library, and Tune, a scalable hyperparameter tuning library.

Ray provides a simple, universal API for building distributed applications. Ray is packaged with the following libraries for accelerating machine lear

23.3k Dec 31, 2022
K-means clustering is a method used for clustering analysis, especially in data mining and statistics.

K Means Algorithm What is K Means This algorithm is an iterative algorithm that partitions the dataset according to their features into K number of pr

1 Nov 01, 2021
Machine Learning toolbox for Humans

Reproducible Experiment Platform (REP) REP is ipython-based environment for conducting data-driven research in a consistent and reproducible way. Main

Yandex 663 Dec 31, 2022
Python library which makes it possible to dynamically mask/anonymize data using JSON string or python dict rules in a PySpark environment.

pyspark-anonymizer Python library which makes it possible to dynamically mask/anonymize data using JSON string or python dict rules in a PySpark envir

6 Jun 30, 2022
Reproducibility and Replicability of Web Measurement Studies

Reproducibility and Replicability of Web Measurement Studies This repository holds additional material to the paper "Reproducibility and Replicability

6 Dec 31, 2022
A Python implementation of GRAIL, a generic framework to learn compact time series representations.

GRAIL A Python implementation of GRAIL, a generic framework to learn compact time series representations. Requirements Python 3.6+ numpy scipy tslearn

3 Nov 24, 2021
GRaNDPapA: Generator of Rad Names from Decent Paper Acronyms

Generator of Rad Names from Decent Paper Acronyms

264 Nov 08, 2022
Microsoft contributing libraries, tools, recipes, sample codes and workshop contents for machine learning & deep learning.

Microsoft contributing libraries, tools, recipes, sample codes and workshop contents for machine learning & deep learning.

Microsoft 366 Jan 03, 2023
Magenta: Music and Art Generation with Machine Intelligence

Magenta is a research project exploring the role of machine learning in the process of creating art and music. Primarily this involves developing new

Magenta 18.1k Dec 30, 2022
Forecast dynamically at scale with this unique package. pip install scalecast

🌄 Scalecast: Dynamic Forecasting at Scale About This package uses a scaleable forecasting approach in Python with common scikit-learn and statsmodels

Michael Keith 158 Jan 03, 2023
Scikit-Garden or skgarden is a garden for Scikit-Learn compatible decision trees and forests.

Scikit-Garden or skgarden (pronounced as skarden) is a garden for Scikit-Learn compatible decision trees and forests.

260 Dec 21, 2022
Accelerating model creation and evaluation.

EmeraldML A machine learning library for streamlining the process of (1) cleaning and splitting data, (2) training, optimizing, and testing various mo

Yusuf 0 Dec 06, 2021
This is an auto-ML tool specialized in detecting of outliers

Auto-ML tool specialized in detecting of outliers Description This tool will allows you, with a Dash visualization, to compare 10 models of machine le

1 Nov 03, 2021
PLUR is a collection of source code datasets suitable for graph-based machine learning.

PLUR (Programming-Language Understanding and Repair) is a collection of source code datasets suitable for graph-based machine learning. We provide scripts for downloading, processing, and loading the

Google Research 76 Nov 25, 2022
Credit Card Fraud Detection, used the credit card fraud dataset from Kaggle

Credit Card Fraud Detection, used the credit card fraud dataset from Kaggle

Sean Zahller 1 Feb 04, 2022
This is a curated list of medical data for machine learning

Medical Data for Machine Learning This is a curated list of medical data for machine learning. This list is provided for informational purposes only,

Andrew L. Beam 5.4k Dec 26, 2022
High performance implementation of Extreme Learning Machines (fast randomized neural networks).

High Performance toolbox for Extreme Learning Machines. Extreme learning machines (ELM) are a particular kind of Artificial Neural Networks, which sol

Anton Akusok 174 Dec 07, 2022
Price Prediction model is used to develop an LSTM model to predict the future market price of Bitcoin and Ethereum.

Price Prediction model is used to develop an LSTM model to predict the future market price of Bitcoin and Ethereum.

2 Jun 14, 2022
CobraML: Completely Customizable A python ML library designed to give the end user full control

CobraML: Completely Customizable What is it? CobraML is a python library built on both numpy and numba. Unlike other ML libraries CobraML gives the us

Sriram Govindan 14 Dec 19, 2021