Training machines using unbiased data and methodology is something that should be considered when designing artificial intelligence (AI) systems. Machine decisions can affect our rights, and we need to ensure that AI does not absorb biases by being trained on biased data. Researchers at the University of Bristol have investigated potential biases and looked at ways in which they could be removed.
The creation of a new generation of AI systems that can be trusted to make fair and unbiased decisions is an urgent task for researchers. As AI rapidly conquers technical challenges related to predictive performance, we are discovering a new dimension to the design of such systems that must be addressed: the fairness and trust in the system’s decisions.
The research team at Bristol addressed this critical issue of trust in AI by not only proposing a new high standard for models to meet (being agnostic to a protected concept) but also proposing a way to achieve such models. The model was defined to be agnostic, with respect to a set of concepts, if it was shown that it makes its decisions without ever using these concepts. This is a much stronger requirement than in distributional matching or other definitions of fairness. The team focussed on the case where a small set of contextual concepts should not be used in decisions, and can be exemplified by samples of data. They demonstrated how ideas developed in the context of domain adaptation can deliver agnostic representations that are important to ensure fairness, and therefore, trust.
The team used a Domain Adversarial Neural Network (DANN) method. Their experiments demonstrated that this method can successfully remove unwanted contextual information, and makes decisions for the right reasons. DANNs are a type of Convolutional Neural Network (CNN) that can achieve an agnostic representation using three components: 1) a feature extractor 2) a label prediction output layer and 3) an additional protected concept prediction layer.
A technically different but analogous process led the same team to explore unbiased representations in natural language processing, demonstrating that it is possible to remove gender bias from the way in which we represent words – an essential step if we want our algorithms to screen CVs and resumes.
While demonstrated in this work by ignoring the physical background context of an object in an image, the same approach could be used to ensure that other contextual information does not make its way into black-box classifiers deployed to make decisions about people in other domains and classification tasks.
Read the published book chapters to find out more:
Machine Decisions and Human Consequences Teresa Scantamburlo, Andrew Charlesworth, Nello Cristianini. Also published as a chapter in: Algorithmic Regulation, Oxford University Press (2019).
Right for the Right Reason: Training Agnostic Networks Sen Jia, Thomas Lansdall-Welfare and Nello Cristianini. Also published in: Advances in Intelligent Data Analysis XVII, Lecture Notes in Computer Science, vol 11191. Springer (2018).
Biased Embeddings from Wild Data: Measuring, Understanding and Removing Adam Sutton, Thomas Lansdall-Welfare and Nello Cristianini. Also published in: Advances in Intelligent Data Analysis XVII, Lecture Notes in Computer Science, vol 11191. Springer (2018).
This work is part of the ERC ThinkBIG project, Principal Investigator Nello Cristianini, University of Bristol.