At Rise of AI 2023 in Berlin, Shaun Moore, Chief AI Officer at Pangiam and former founder of facial recognition company TrueFace, which Pangiam acquired in 2021, argues that facial recognition can be deployed responsibly only if "humanity first" guides every stage of development and deployment. He traces the technology's history from Woodrow Wilson Bledsoe's 1960s hand-geometry comparisons through the first correct computer match in 1991, its 2001 Super Bowl use on 60,000 to 70,000 people, Facebook's 2010 mass-market photo tagging, and the 2017 iPhone X, and notes there are now about 1 billion cameras worldwide. He identifies four points where bias enters facial recognition, data collection, algorithmic design, feature selection and deployment, and lists specific bias types including sampling, labeling, historical, representation and algorithmic bias, illustrating algorithmic bias with a model that matched an albino wallaby to a cat based on color. He describes measuring bias with the FairFace dataset and a Department of Homeland Security study finding algorithms learn gender and race features alongside identity, and presents his company's SMU bias lab as a step toward explainability and industry-wide bias certification.
Key takeaways
- Moore says facial recognition template data is roughly two kilobytes per person, about 500 to 600 numerical values, never the original image.
- He identifies four stages where bias enters facial recognition systems: data collection, algorithmic design, feature selection and deployment.
- A Department of Homeland Security study he cites found algorithms learn gender and race features alongside identity, and removing 10 percent of ethnicity-linked features raises fairness but lowers accuracy.
- He says synthetic data is not yet accurate enough for training facial recognition because generated faces tend to be unrealistically symmetrical.
- He argues there is still no generally accepted industry or federal definition of bias in AI, which he considers a necessary next step.
Detailed notes
From hand-drawn geometry to a billion cameras
Moore traces facial recognition's history from Woodrow Wilson Bledsoe's 1960s work, done jointly with a US intelligence agency, comparing hand-drawn facial geometry, to the first correct computer-run match in 1991 out of about 2,000 images. He describes its first large-scale commercial use scanning 60,000 to 70,000 people at a Super Bowl, later shut down over privacy concerns, followed by Facebook's 2010 rollout of automatic photo tagging and the 2017 iPhone X, which he calls the turning point that put facial recognition directly in consumers' hands. He notes there are now roughly 1 billion cameras worldwide, that Raspberry Pi-sized devices can store 100,000 to 1 million faces and run matching fully offline, and that deep learning, introduced to the field in 2015, lets his company's systems claim to scan a billion faces, roughly the world's population, in under ten seconds. He cites NIST testing across more than 350 companies' algorithms, reporting accuracy above 99 percent even at a 1-to-12-million identification scale.
How facial recognition works and where bias enters
Moore explains that systems detect facial landmarks, normalize the image and extract features into a numerical template of about 500 to 600 values, roughly two kilobytes, rather than storing images, and run either one-to-many searches or one-to-one match verification. He identifies four stages where bias can enter: data collection, algorithmic design, feature selection and deployment. Sampling bias occurs when training data does not match the real-world application, such as a model trained for CCTV used in access control. Labeling bias comes from human annotators' own prejudices. Algorithmic design bias can arise from processing sensitive features like gender or eye-area geometry, and overfitting occurs when training data is concentrated on one gender or ethnicity. Feature selection can introduce omitted-variable bias by excluding relevant facial information, and deployment bias appears when the operating context differs from the training context, raising false positives and lowering recognition rates.
Why bias exists: cost, data and cameras
Moore attributes persistent bias partly to economics: commercially available training data is cheap but typically concentrated in one ethnic group, and collecting data at scale is expensive, citing Google's reported five-dollar-per-identity payments in Times Square to build training sets of tens of millions of faces. He lists further causes including country-based data gaps when training sets are not globally representative, flawed or blurry images, unrepresentative or incomplete datasets, and camera quality and lighting, which he says a recent Department of Homeland Security technology rally showed materially affects results even when algorithms are blamed. He also names historical, representation, measurement, evaluation, aggregation, population and algorithmic bias as recognized categories the industry has to test for.
Quantifying and explaining bias
Moore illustrates algorithmic bias with an example where a deep learning model matched an image of an albino wallaby to a cat because it had learned to focus on color rather than relevant facial features. He describes FairFace, an independent dataset of about 70,000 faces across seven major ethnicity categories, as essentially the only current resource for benchmarking bias, and shows his own company's year-over-year accuracy gains at a fixed operating threshold after using such evaluation to target data collection. He cites a Department of Homeland Security paper finding that facial recognition algorithms learn features clustered around gender and race, not just individual uniqueness, and that removing 10 percent of ethnicity-related features reduces accuracy but increases fairness, framing this trade-off as an unresolved question for vendors and buyers alike.
Mitigation: the SMU bias lab and policy questions
Moore describes launching the Intelligence Systems and Bias Examination Lab ("Isabel") with Southern Methodist University in Dallas, aiming to bring peer-reviewed research standards to bias investigation, with explainability as the key goal so mitigation is no longer based on assumptions. He raises federally curated, sandboxed datasets as one possible way to reach parity across ethnicities without vendors owning the data, and says synthetic data is not yet accurate enough for training because generated faces tend to be too symmetrical compared with real human faces. He calls for an industry-wide, ideally global or federal, definition of bias and a standard way to measure it, alongside self-hosted deployment options that give customers ownership and transparency over their data, and closes by reiterating that "humanity first" must be a non-negotiable principle guiding investors, researchers, designers and customer support alike.
Quotes
Humanity first has got to be a non-negotiable for the design, development, deployment of this technology.
Shaun Moore
We can scan through a billion faces a second, the population of the Earth, in under ten seconds.
Shaun Moore
Questions answered in this talk
What are the main sources of bias in facial recognition technology?
According to Shaun Moore at Rise of AI 2023, bias can enter through data collection (sampling and labeling bias), algorithmic design, feature selection, and deployment in a context different from the one the system was trained for.
How is bias measured in facial recognition algorithms?
According to Shaun Moore at Rise of AI 2023, his company benchmarks algorithms against FairFace, an independent dataset of about 70,000 faces across seven major ethnicity categories, and compares similarity scores at a fixed operating threshold to track year-over-year accuracy gains.
Is synthetic data ready for training facial recognition systems?
According to Shaun Moore at Rise of AI 2023, synthetic data is not yet accurate enough because generated faces tend to be more symmetrical than real human faces, which reduces performance.