Debiasing Convolutional Neural Networks via Meta Orthogonalization

Abstract

While deep learning models often achieve strong task performance, their successes are hampered by their inability to disentangle spurious correlations from causative factors, such as when they use protected attributes (e.g., race, gender, etc.) to make decisions. In this work, we tackle the problem of debiasing convolutional neural networks (CNNs) in such instances. Building off of existing work on debiasing word embeddings and model interpretability, our Meta Orthogonalization method encourages the CNN representations of different concepts (e.g., gender and class labels) to be orthogonal to one another in activation space while maintaining strong downstream task performance. Through a variety of experiments, we systematically test our method and demonstrate that it significantly mitigates model bias and is competitive against current adversarial debiasing methods.

arXiv commentAccepted to NeuRIPS 2020 Workshop on Algorithmic Fairness through the Lens of Causality and Interpretability (AFCI). Supplemental materials provided at: https://drive.google.com/drive/folders/1klIAqZDgg3sCVmzFjLw5Y_T-GTc2E3oh?usp=sharing

References in corpus (13)

• Distributed Representations of Words and Phrases and their Compositionality
• Model-Agnostic Meta-Learning for Fast Adaptation of Deep Networks
• Semantics derived automatically from language corpora contain human-like biases
• Equality of Opportunity in Supervised Learning
• Explaining NonLinear Classification Decisions with Deep Taylor Decomposition
• Man is to Computer Programmer as Woman is to Homemaker? Debiasing Word Embeddings
• SmoothGrad: removing noise by adding noise
• Data Decisions and Theoretical Implications when Adversarially Learning Fair Representations
• Concept Whitening for Interpretable Image Recognition
• Achieving Fairness through Adversarial Learning: an Application to Recidivism Prediction
• Women also Snowboard: Overcoming Bias in Captioning Models
• Benchmarking Attribution Methods with Relative Feature Importance
• Gender-preserving Debiasing for Pre-trained Word Embeddings