References
— Goodfellow, I., Bengio, Y., Courville, A. "Deep Learning", MIT Press, 2016.
— Henderson, P., Islam, R., Bachman, P., Pineau, J., Precup, D., Meger, D. "Deep Reinforcement Learning that Matters", arXiv:1709.06560, 2017
— Russakovsky et al. "ImageNet Large Scale Visual Recognition Challenge", International Journal of Computer Vision, 2015
— Russell, S.J., Norvig, P. "Artificial Intelligence: A Modern Approach", 3rd edition, Prentice Hall, 2009.
13 The IEEE P7003 Standard for Algorithmic Bias Considerations
Ansgar Koene
University of Nottingham
13.1 Statement
The rapid advance in the application of algorithmic decision making and machine learning methods to real-world applications, like screening of job applicate CVs, public-sector resource allocation (e.g. policing) and autonomous vehicles, with potentially significant impact on peoples’ lives has generated an urgent need for practical guidelines and industry (self-)regulation in order to ensure that the highest standards of responsible conduct are applied as these powerful new algorithmic systems are developed and deployed.
A key challenge when it comes to the regulation of algorithmic decision making systems is that any evaluation of the bias/fairness of these system must take into account the inherently socio-technical context of how the system is (intended to be) used. When used for impactful decisions, the norms that an algorithmic system must obey are not just statistical, but also legal, moral and cultural [Dansk and London 2017].
In recognition of these challenges professional associations such as the ACM and the FAT/ML community have responded by publishing Principles for Algorithmic Accountability [ACM 2017, FATML] and a Social Impact assessment statement for Algorithms [FATML]. Around the same time the IEEE launched the IEEE Global Initiative on Ethics of Autonomous and Intelligence Systems which is developing a document [IEEE 2017] and a series of ethics based industry standards aimed at moving the discussion beyond statements of principles toward practical standards and policies.
13.2 Future challenges
- There is a need for more multidisciplinary coordinated thinking about the ways in which algorithms impact individuals and society.
- There is a need for clear assessment and certification regimes to communicate to users which algorithmic systems have implemented best practices for avoiding algorithmic bias.
- There is a need for research on effective benchmarking and impact assessment methods, especially regarding social impacts that go beyond statistical assessment of disparate outcomes.