Our company ethos revolves around increasing diversity and representation for POC and there is no greater under-represented sector than machine learning and AI
Building Inclusive Transparent Explainable Accurate Robust AI
Publicly available datasets are heavily Caucasian biased in their weighting of ethnicities and facial diversity.
Common facial biometric applications are poorly trained for darker skin tones, features and low contrast scenarios.
”Melanoma is most common in white skin. Black people are less likely to get it, but they are more likely to die from it.Poppy Noor,The Guardian, New York
Why Racial & Sex Fairness
Matters In Machine Learning
AI Is Used To…
- Conduct demographic surveys of racial groups for census counts (1)
- Measure racial diversity in geographic regions and how racial groups interact amongst each other (2)
- Analyse facial traits for facial recognition and biometrics (3)
- Facial recognition is poor at guessing age in people of colour (4)
- Facial recognition struggles with diversity in racial groups, especially in terms of headwear (veils, hijabs), facial hair (moustaches, beards) and facial ornaments such as glasses
- Facial recognition tools produce higher false positives for African and East Asian faces compared to Caucasian (6)
- Enforce the law through photo matching CCTV to facial databases
- High false positives can lead to the wrong person being convicted (7)
Cosmetic Concern Tool V1.2
Racial Fairness Dataset V1.54
Cephalograph Marker V1.0
Skincare Assessment Tool V2.0
Human Face GAN Project
Our AI Goals
Improve Racial Parity & Fairness In AI
Did you know that only 20% of faces in datasets are black? We tested this for ourselves using StyleGAN to find that… See the results for yourself here.
Increase Low Contrast & Low Light Detection In Computer Vision
Darker skinned users are often neglected by facial recognition algorithms as the barrier between their skin and low light backgrounds become difficult for the AI to detect. We aim to improve this through a number of image processing techniques in our products.
Improve Transparency & Explainability In Predictions
AI predictions are only as useful as they are explainable. Transparent and explainable AI predictions are more likely to be trusted and be adopted into mainstream use. One of our key challenges is translating technical data science into user friendly outputs.