Space Bias

Space bias, or the uneven representation of different groups or features within a model's learned space, is a significant problem across machine learning domains, hindering fairness and accuracy. Current research focuses on identifying and mitigating these biases in various models, including large language models (LLMs) and computer vision systems, using techniques like activation steering, adversarial training, and data augmentation strategies designed to improve feature representation and balance class distributions. Understanding and addressing space bias is crucial for developing reliable and equitable AI systems, impacting fields ranging from natural language processing and recommendation systems to medical diagnosis and autonomous driving.

Papers

February 1, 2024

Investigating Bias Representations in Llama 2 Chat via Activation Steering
Dawn Lu, Nina Rimsky
Gender Bias Societal Bias LLaMa LlamaCare E Chat Space Bias Activation Steering

December 10, 2023

Exploiting Representation Bias for Data Distillation in Abstractive Text Summarization
Yash Kumar Atri, Vikram Goyal, Tanmoy Chakraborty
Representation Learning Abstractive Text Summarization Data Distillation Semantic Encoder Space Bias

September 29, 2023

FACTS: First Amplify Correlations and Then Slice to Discover Bias
Sriram Yenamandra, Pratik Ramesh, Viraj Prabhu, Judy Hoffman
Model Slicing Slice by Slice Bias Related Issue Bias Discovery Space Bias Correlation Bias

September 27, 2023

Automatic Feature Fairness in Recommendation via Adversaries
Hengchang Hu, Yiming Cao, Zhankui He, Samson Tan, Min-Yen Kan
Adversarial Perturbation Effective Recommendation Adversary Agent Fairness Aware Generalizable Feature Space Bias Imbalanced Training

August 24, 2023

Uniformly Distributed Category Prototype-Guided Vision-Language Framework for Long-Tail Recognition
Siming Fu, Xiaoxuan He, Xinpeng Ding, Yuchen Cao, Hualiang Wang
Pre Trained Vision Language Model Long Tailed Recognition Long Tailed Learning Category Shift Long Tail Recognition Space Bias

August 19, 2023

Partition-and-Debias: Agnostic Biases Mitigation via A Mixture of Biases-Specific Experts
Jiaxuan Li, Duc Minh Vo, Hideki Nakayama
Mixture Component Bias Mitigation Bias Removal Bias Reduction Space Bias Based Debiasing

May 30, 2023

Fighting Bias with Bias: Promoting Model Robustness by Amplifying Dataset Biases
Yuval Reif, Roy Schwartz
Absolute Stance Bias NLP Model Model Robustness Dataset Bias Common Bias Space Bias Addressing Bias

May 24, 2023

ReLU Neural Networks with Linear Layers are Biased Towards Single- and Multi-Index Models
Suzanna Parkinson, Greg Ongie, Rebecca Willett
Neural Network Neural Network Architecture Single Label Shallow ReLU ReLU Neural Network Layer Representation Linear Layer Space Bias Multi Index Model

May 22, 2023

Measuring Inductive Biases of In-Context Learning with Underspecified Demonstrations
Chenglei Si, Dan Friedman, Nitish Joshi, Shi Feng, Danqi Chen, He He
Large Language Model Context Learning Inductive Bias Suboptimal Demonstration Space Bias

November 22, 2022

Supervised Contrastive Learning on Blended Images for Long-tailed Recognition
Minki Jeong, Changick Kim
Contrastive Learning Data Augmentation Supervised Contrastive Learning Long Tailed Recognition Space Bias Blended Image

August 24, 2022

Identifying and Overcoming Transformation Bias in Forecasting Models
Sushant More
Forecasting Model Bias Correction Space Bias Tweedie Distribution

February 2, 2022

The Role of Linear Layers in Nonlinear Interpolating Networks
Greg Ongie, Rebecca Willett
Neural Network Integral Role Neural Network Architecture Implicit Bias Optimal Representation Linear Layer Space Bias Neural Interpolation

Space Bias

Papers

Investigating Bias Representations in Llama 2 Chat via Activation Steering

Exploiting Representation Bias for Data Distillation in Abstractive Text Summarization

FACTS: First Amplify Correlations and Then Slice to Discover Bias

Automatic Feature Fairness in Recommendation via Adversaries

Uniformly Distributed Category Prototype-Guided Vision-Language Framework for Long-Tail Recognition

Partition-and-Debias: Agnostic Biases Mitigation via A Mixture of Biases-Specific Experts

Fighting Bias with Bias: Promoting Model Robustness by Amplifying Dataset Biases

ReLU Neural Networks with Linear Layers are Biased Towards Single- and Multi-Index Models

Measuring Inductive Biases of In-Context Learning with Underspecified Demonstrations

Supervised Contrastive Learning on Blended Images for Long-tailed Recognition

Identifying and Overcoming Transformation Bias in Forecasting Models

The Role of Linear Layers in Nonlinear Interpolating Networks