Machine learning (ML) is the future, and its development and implementation is pivotal to the success of technology companies today. Yet the way Artificial Intelligence (AI) is being currently implemented by many companies, does not include diversity as an essential ingredient to their success. Yet diversity is a critical if not mandatory requirement in developing versatile, holistic, and impactful uses cases for AI and ML.
Diversity deficit in the AI sector
The ‘AI Now’ Institute of New York University, conducted a survey and noted only 15% of all AI researchers at Facebook and 10% of all AI researchers at Google are women. What’s more, less than 25%1 of PhDs were awarded to women and minorities in the fields of computer science.
Unsurprisingly, men hold 79%2 of the leadership roles within the sector. This male dominated hierarchy can be seen as problematic, resulting in bias towards how teams analyse information and how their outcomes are impacted.
AI dependence on data
Before I delve into this topic, I want to first touch upon ‘saliency datasets’. They are basically large chunks of data preprocessed by humans that can be used to train artificial intelligence. Let’s say we want to develop an algorithm that can detect faces in each photo. We would need to gather pictures of people and label their faces — this process yields us a ‘saliency dataset’. This task is undeniably unfeasible for a tech company to manage, any programmer would perhaps sample an MIT saliency dataset, as it appears to be a credible source. However, upon closer analysis of the dataset, we can see the photos contain more pictures of cats than people of diverse ethnicity, leading to the algorithm being objectively worse at recognising the face of a varied population and hence challenging the very accuracy of the analysis.
Therein lies the crux of the issue, may be unconsciously, we are attempting to design systems that embody egalitarianism, while using data that is riddled with innate biases! In some sense of the idea, we are unintentionally degrading the quality of the analysis and promoting involuntarily biases by utilising fundamentally homogenous data. Unless a paradigm shift happens toward collecting and utilising data with a higher variance, the quality of the analysis and the accuracy of the inferences will remain highly subjective and highly prejudiced.
Interestingly, we have had a similar problem in the past. Research papers published in the late 19th century and early 20th century were often built upon biased data collection and analysis. Especially in the case of medical research, all of the information was centered around the treatment of just the Anglo-Saxon male population. By increasing regulatory standards, and introducing the concept of peer-reviewed papers, we were able to conduct better research and therefore better and more accurate diagnosis and treatments that represent the diversity of the real world.
In today’s digital economy, we have all experienced Google and Facebook fight vehemently to keep their trademark bots a secret. For this technology to mature, we need to continue to challenge and demand transparency — about how the technology that guides and recommends our choices and decision making are built. This will help us ensure higher quality and accuracy of our decision making in our data-driven world.
Forging past a single definition of diversity for AI / ML
Diversity is often seen in the content we consume, where casts consist of people of different races, sexualities, and beliefs. This makes us view diversity in a very monochromatic way of merely hiring and including people different from ourselves. However, it is more about how different perspectives can enhance the way we approach problems. We all want more women in technology but forcing diversity from a mainstream perspective breeds discord. People may be reduced simply to the ‘diversity hire of the month,, rather than a meaningful inclusion to the team. Inclusivity is more about encouraging previously dis-enfranchised groups to participate in fields they were historically restricted to participate in and not just a metric to track gender inclusivity.
Companies are increasingly using data to make decisions and draw insights, which in turn leads to better products or services and higher levels of innovation. The challenge is that the products being developed are unilaterally being catered to a specific group of people, due in part because of our homogenous data. This can lead to poor product design and underrepresented features, which negatively impact a company’s brand image as well as its bottom line. An example close to home is our virtual assistants — Siri or Alexa. They are far worse at comprehending all types of accents. This leads to many people feeling excluded and dis-enfranchised. Not a good product development strategy when more than 80 % of the market they serve are global consumers with multiple dialects.
A study from Harvard recently indicated companies that embark on diversity programs innovate and outperform others. Diversity unlocks product innovation and drives the market growth. It increases the chances of understanding customers and their values, by providing a different point of view which result in creative breakthroughs.
So, if you’re serious about building an AI product that works globally and not just for a specific market, then you need diversity in the team. Diversity inspires an inclusive work culture and helps develop intense empathy and understanding within your technology, product design and creation activities. This will not only help shape your workplace into adopting a more inclusive and accepting culture, but it will accelerate product quality and design and its acceptance and reach in the global markets. More perspectives mean more varying viewpoints, leading to a more richer and appealing product design and quality, impacting a far larger potential customer base.
AI and ML are the key principles that will drive the future of digitally powered economies. Technology companies that are keen to harness the opportunity this presents need to judiciously integrate diversity in all aspects of their business processes and people in order to succeed in innovating world class products that will be globally inclusive and impactful.
