Turning a Blind Eye: Explicit Removal of Biases and Variation from Deep Neural Network Embeddings (1809.02169v2)

Abstract: Neural networks achieve the state-of-the-art in image classification tasks. However, they can encode spurious variations or biases that may be present in the training data. For example, training an age predictor on a dataset that is not balanced for gender can lead to gender biased predicitons (e.g. wrongly predicting that males are older if only elderly males are in the training set). We present two distinct contributions: 1) An algorithm that can remove multiple sources of variation from the feature representation of a network. We demonstrate that this algorithm can be used to remove biases from the feature representation, and thereby improve classification accuracies, when training networks on extremely biased datasets. 2) An ancestral origin database of 14,000 images of individuals from East Asia, the Indian subcontinent, sub-Saharan Africa, and Western Europe. We demonstrate on this dataset, for a number of facial attribute classification tasks, that we are able to remove racial biases from the network feature representation.

• The paper introduces the JLU algorithm, which uses a joint loss function to explicitly remove bias from DNN embeddings.
• It leverages the LAOFIW dataset of 14,000 diverse images to quantitatively assess bias removal in image classification tasks.
• Experiments demonstrate up to 20% improvement and reduced KL divergence, confirming the method's effectiveness in unlearning bias.

Bias Mitigation in Deep Neural Networks: An Evaluation of JLU in Image Classification

The paper "Turning a Blind Eye: Explicit Removal of Biases and Variation from Deep Neural Network Embeddings" examines the challenge of biases in deep neural network (DNN) embeddings used for image classification. The authors introduce the Joint Learning and Unlearning (JLU) algorithm, a novel approach aimed at ensuring that DNN models do not learn and internalize unwanted biases from the training dataset. This paper addresses significant issues associated with dataset biases—such as gender and ancestral origin biases—and demonstrates the effectiveness of the algorithm through extensive experiments.

Key Contributions

The paper outlines two primary contributions:

1. Algorithm for Bias Removal: The authors propose a supervised-learning algorithm that facilitates learning a feature representation invariant to multiple spurious variations. The algorithm is inspired by domain adaptation methods and integrates a confusion loss to ensure that the trained model becomes agnostic to specified biases.
2. Ancestral Origin Dataset: Included in the research is the introduction of the "Labeled Ancestral Origin Faces in the Wild (LAOFIW)" dataset, which consists of 14,000 images representing diverse ancestral origins. This dataset serves both as an experimental testbed and a tool for mitigating biases related to racial features in DNNs.

Methodology

The core innovation in the paper is the JLU algorithm, which employs a joint loss function over primary and secondary datasets. The primary dataset focuses on classification tasks, while secondary datasets address spurious variations. The methodology emphasizes the creation of a feature representation ($\theta_{repr}$) capable of distinguishing primary tasks while simultaneously being indifferent to biases such as gender and race. This is achieved through the implementation of a confusion loss that encourages a uniform distribution across specified secondary tasks, progressively reducing the network's sensitivity to these biases.

Experiments

Three key experiments encapsulate the efficacy of the JLU algorithm:

1. Removal of Bias from Network: Applied to a gender-agnostic age classification task using a gender-biased dataset, the experiment demonstrated a marked decrease in classification discrepancies between genders. A notable result is a reduction in Kullback-Leibler divergence between age prediction distributions for men and women, indicating effective bias removal.
2. Extreme Bias Mitigation: Gender classifiers were trained on datasets exhibiting extreme age biases. The JLU algorithm improved classification performance by up to 20% compared to traditional baselines, showcasing its robustness against extreme bias scenarios.
3. Simultaneous Bias Removal: The ability of JLU to unlearn multiple spurious variations concurrently was tested. Improvements in primary task accuracy were observed along with substantial reductions in secondary classification accuracies, nearly to the level of random chance, confirming effective unlearning of biases.

Implications and Future Directions

The implications of this research are substantive, particularly in contexts where fairness and transparency are paramount. By providing a mechanism to ensure that predictive models do not base decisions on biased representations, the work stands to substantially improve the integrity of DNN applicability in sensitive domains such as government policy, healthcare, and employment.

Future research could explore dynamic weighting of spurious variation classifiers during the training process, addressing the differential difficulty in removing certain biases. Additionally, the extension of JLU to other types of neural networks and diverse data modalities represents a promising direction for future studies.

In conclusion, the paper effectively elucidates a methodological framework to address bias in neural networks, with the JLU algorithm presenting a calculated approach toward fairer and more reliable AI systems. Through rigorous experimentation, it sets a foundational benchmark for future endeavors aiming to tackle inherent biases in AI models.