Bias in Facial Recognition
This post was originally published as a recipe on https://ai.science. See this original recipe by Willie Costello for an interactive version of this post and to comment or collaborate.
Increased utilization of algorithmic technologies like Artificial Intelligence (AI) and Machine Learning (ML) is partly driven by a desire to avoid the errors and biases that plague human’s treatment and interpretation of data. However, the real-world performance of current AI and ML technologies fall short of this ideal¹. In this brief report, I discuss a handful of applications of these technologies which demonstrate ways in which they can mirror and amplify human biases rather than avoid them.
A particularly salient example of this comes from the algorithmic instantiation of stereotypic beliefs associating people of color with criminality. Despite deployments in many states, criminal recidivism prediction tools have misclassified Black defendants as being twice as likely for violent recidivism relative to White defendants, while White recidivists (people that actually re-offended) were misclassified as “low risk” 63.2 percent more often than Black defendants². The injuries caused by these inaccuracies are real³ and highlight what the people and communities threatened by this technology already know: Technologies constructed by the (sometimes well-intentioned) oppressing class tend to reinforce existing systems of oppression rather than dismantle them⁴.
Race-related biases also plague a widely implemented class of Facial Recognition Technologies (FRT) used to identify and process facial characteristics when they appear in video or image recordings⁵. FRTs fall into a few broad categories: Face detection algorithms identify the presence of a face or faces within an input stream. Face attribute classification algorithms attempt to describe either objective (e.g., smiling or frowning) or abstract facial characteristics (e.g., attractiveness or age) of detected faces. Face Verification and Face Identification algorithms attempt to discern the actual identity of people in an input stream. The accuracy of a FRT is heavily influenced by (at least) two interrelated features:
- The diversity of facial characteristics used to train the algorithm, and
- The standards of accuracy used to assess the algorithm’s accuracy.
With respect to training, each of the above classes of FRTs suffer from being developed and trained on data sets made up of primarily White and Male facial characteristics. As a consequence, people whose facial characteristics do not match these demographic categories can be invisible to Face Detection⁶ , misunderstood by Face attribute classification⁷, and misidentified by Face Verification and Identification softwares⁷ .
With respect to accuracy, applications of FRT are often assessed by the proportion of errors made across a long series of trials. This benchmarking often occurs in a context that is less noisy than the intended application and by utilizing test data that are similarly White and Male-centric as the training data. As a consequence, the predictions of the accuracy of these systems if often inflated and the types of errors outlined above remain invisible.
What can be done to begin rectifying these issues? A first step is to raise awareness about these issues. More actionable steps include increasing the phenotypic and demographic diversity of individuals represented in the data sets used to train and evaluate FRTs. Moreover, it would be helpful to develop international standards for the training, operation, application, and evaluation of FRTs.
If you are interested in learning more about Face Recognition Technologies, the biases that exist in these and related systems, or are simply interested in learning more about social and industrial applications of artificial intelligence and machine learning, consider following the superscripts above as well as engaging with the original recipe associated with this blog. This post is intended as a gateway to the work of others dedicating their professional and personal lives to combatting and solving issues like these.
Aggregate Intellect
Aggregate Intellect is a Global Marketplace where ML Developers Connect, Collaborate, and Build. Connect with peers & experts at https://ai.science or Join our Slack Community.
- Check out the user generated Recipes that provide step by step, and bite sized guides on how to do various tasks
- Join our ML Product Challenges to build AI-based products for a chance to win cash prizes
- Connect with peers & experts through the ML Discussion Groups or Expert Office Hours
Contact Brandon Terrizzi, PhD:
If you enjoyed this post and/or have questions or comments, feel free to add them below. If you are interested in adding me to your professional network or connecting in another way, consider navigating to my LinkedIn and ResearchGate pages.
