Outsmarting AI beauty bias, Part Two: Adapting to rapidly evolving landscape

14 Dec 2023 --- A greater dependence on AI for boosting personalization and diagnosis has spurred the need to further understand how the algorithms work.

In the final part of this series, Personal Care Insights learns how EveLab Insight and Haut.AI establish unbiased AI models, collect data and keep their technologies up to date as beauty swiftly evolves. 

Recently, a research paper titled “How Artificial Intelligence Adopts Human Biases: The Case of Cosmetic Skincare Industry” looked into the challenges posed by biases in AI development throughout the industry. 

It warned that despite pursuing optimal supply chains, high-quality products and personalized customer experiences, digital technologies come with risks.

We previously explored solutions for overcoming biases with a focus on skin health, including how Haut.AI and EveLab Insight establish a solid basis in objective facts over beauty ideals.

Ensuring diverse data collection
Nick Howard, global strategy director at EveLab Insight, admits managing potential biases can be challenging and shares EveLab Insight’s measures to establish unbiased AI models. They include:

• Collecting diverse and representative data: We strive to gather samples from diverse populations worldwide, encompassing various ages, genders and skin characteristics. For example, when developing a skin type recognition algorithm, our data includes samples representing different skin types, ranging from dry to oily, various degrees of acne and skin issues like pigmentation. We ensure coverage of diverse skin tones and ethnic features.
• Using rich training data from varied scenarios: Our models are trained on data from various scenarios, preventing the training set from reflecting only a single environment or condition. For instance, when developing an algorithm to assess skin texture, we incorporate facial images captured in different environments (such as sunlight and indoor lighting) and account for seasonal variations that impact skin conditions.
• Maintaining accuracy in manual annotations: We involve dermatology experts to establish data annotation standards, ensuring accurate labels for each image. Dermatology specialists categorize different skin conditions in detail, create skin grading charts, and conduct third-party audits to avoid subjective errors in annotation results.

“Nipping bias in the bud”
Comprehensive dataset composition is fundamental to developing AI models, says Georgievskaya.According to Anastasia Georgievskaya, CEO at Haut.AI, dataset composition is fundamental to developing AI models. “The more comprehensive data we collect, the more robust and fair models we develop. At Haut.AI, we maintain high standards of data acquisition.”

“When developing algorithms and AI solutions for skin analysis, a wide range of specialists are involved, including skin physiology and biophysics experts. We aim to ensure that our algorithms accurately reflect real skin metrics, regardless of gender, ethnicity or skin color. Before developing the model, we put significant effort into defining the targets. We identify sensitive parameters, which are unique characteristics of different skin types that manifest in their way.” 

For instance, hyperpigmentation in darker skin will have specific features, underscores Georgievskaya. “We know that such skin is more prone to hyperpigmentation, and certain skin conditions are more prevalent among such individuals. After predicting parameters like gender or ethnicity, we formulate recommendations for compiling the dataset, considering the distribution of different ethnicities and genders.”

“Considering the end users of your models is crucial for designing and developing effective solutions. It is important to ensure that the dataset used for the algorithm is aligned with the target population where it will be applied. It should be noted that the representation of different phototypes in the Indian population will differ from the Asian or American population,” she continues.

Further, Georgievskaya spotlights “annotation” as a critical step in the model development lifecycle. 

“The more accurately the target features are annotated to reflect the functional state of the skin, the better the algorithm will perform on any skin type, regardless of ethnicity or gender. However, bias can often creep into the annotation process. It happens especially when a narrow circle of annotators, who may not be specialized in the issue, are involved. This can negatively affect the algorithm’s performance,” she explains.

“At Haut.AI, we have a diverse annotation team that comprises individuals from different social backgrounds and genders and with varied life experiences. Though they come from different backgrounds, what unites them is their biomedical expertise and focus on delivering quality results. To ensure a comprehensive understanding of the diversity of features across different skin types, the annotation team conducts coordination sessions regularly before starting a new project and throughout its duration.”

Updating the system
Howard says EveLab Insight follows dermatological standards over aesthetic standards to evaluate and adapt algorithms according to changes in societal perceptions and beauty standards over time.

“Dermatological standards are based on scientific principles from medicine and biology, focusing on the health and physiological characteristics of the skin. In contrast, aesthetic standards are more subjective and can be influenced by societal culture and individual perspectives,” he says. 

“However, a small portion of our features involves societal perceptions and aesthetic standards, such as our detection tool recommending lipstick shades based on skin tone. To constantly evaluate and adapt to changes, we implement the following three measures:

• Partnerships: We establish partnerships with beauty and skin care brands, as well as professional dermatologists, to co-develop with significant skin care and beauty brands. This allows us to respond more quickly when specific skin colors or facial features become preferred in popular culture.
• Regular user feedback collection: We collect direct feedback on aesthetic preferences through online surveys, social media monitoring, and customer interviews. For instance, if users tend to choose more natural-toned foundations over previously popular vibrant colors, we will update our recommendation algorithms to favor products in line with current user preferences.
• Diverse algorithm testing: We conduct extensive algorithm testing on different demographics and geographical regions to verify effectiveness on a global scale. Before releasing an algorithm, we thoroughly test it on samples from Asia, Africa, Europe and the Americas to ensure specific cultural aesthetic standards do not restrict it.”

At Haut.AI, algorithms are monitored to ensure effectiveness and updates reflect cultural trends, consumer preferences, and beauty standards.EveLab Insight puts effort into defining targets by identifying sensitive parameters, which are unique characteristics of different skin types (image credit: EveLab Insight).

“To achieve this, we keep track of recent publications and trends in the beauty industry, engage with diverse communities and stakeholders and collaborate with experts to understand evolving beauty standards and societal perceptions. We follow ethical guidelines that prioritize fairness, transparency and inclusivity in AI development, including the EU AI Act and other national documents regulating the ethics of AI use,” says Georgievskaya.

“We incorporate detected changes into our algorithms as fast as possible. We also analyze customer interactions with our platform and their feedback to identify areas where we may need to adjust our algorithms or recommendations to meet their needs better.”

EU members of parliament recently reached a deal with the Council on the AI Act for comprehensive rules for trustworthy AI. The rules will establish obligations for AI based on its potential risks and level of impact.

By Venya Patel