Why We Need Women in AI
I got excellent feedback after taking part in the panel discussion, and since it was not recorded I thought I would write a blog post on some of what I discussed, my response to the core questions and some other thoughts I had on the topic.
Artificial Intelligence (AI) is often thought of as being new areas that are currently hard to automate, difficult problems to solve using computers, and ultimately replacing humans jobs. Yet we have been using a form of AI or Machine Learning (ML) since the 1950’s as Artificial neural network, later on adopted by businesses in 1970’s initially as decision support systems, and later evolving into data mining, business intelligence, analytics & insights, and more recently data science. What has changed is that the sales and marketing teams are now involved, and sometimes even overpromising on what is possible! Yet there has also been an increase in computing power, storage capacity with massive datasets collected from a larger number of sources, and open source data science code, packages and tutorials that are readily available. All this has helped it become more accessible to many more organisations and individuals.
Q. The problems of building AI and the biases it will inherit & why it is crucial that AI is built with equality at the core
In terms of individuals, I think you can break this into three areas: the data, experience, and context / domain knowledge, in all of which having diversity in your team helps reduce bias.
I would say that the data is an important factor, no matter who analyses it or uses it to train a model, it is the essential raw material used in natural language processing (NLP) and ML. I would break this down into quantitative (quantified measurements like numbers) and qualitative (think about interpreting text, comments, responses etc.), this is an important distinction as not everything can be measured. There is also inherent bias in the datasets, which could for example be dependent on the time frame used such as the changes in society over time. A recent paper from Stanford University word embeddings quantify 100 year on gender and ethnic stereotypes found that word embeddings from texts of 20th and early 21st century are inherently racist and sexist. For example the NLP models found a strong connection of men and women in associated with certain adjectives and occupations. These could be interpreted as stereotypes when automatically learnt, which could be problematic if used in search ranking, filtering users or recommendation engines for example. Historical social views have changed over time, and looking broader, countries and even communities often have different cultures, values, emotions and convictions which can be reflected in the vocabulary they use. If that text is the input used to train the model, then having people from diverse background helps raise questions on the data, and identify what is missing, incomplete or under-represented.
The next factor is the experience in understanding the data and it’s limitations when used in ML models. With experience you will understand how to split the data correctly between training and test sets, what data preparation is required and which machine learning algorithms to use and parameters to set, and how to evaluate and monitor accuracy. For example in a talk on self-driving cars I attended, they discussed that even with all the video and sensor data gathered during trials, they needed to generate synthetic data for training, as it’s not possible to have all possible driving scenarios available. When faced with incomplete data it’s always important to ask why, understand the limitations, and interpret them correctly.
Context and domain knowledge is important when applying ML within your organisation, vertical or industry sector. You might have heard that “correlation does not imply causation”, e.g. when the sales of ice cream on the beach are up you could wrongly conclude that this leads to more people drowning as the two are correlated, but actually it’s just down to there being more people swimming during the summer time. There are a lot more fun spurious correlations. In case you are wondering, the general approach to finding evidence to support causation is to use AB tests and understanding the context and domain.
Q. The risk of AI being built by one demographic … leading to products that only cater to one market
In the future we are starting to look at AI liability, transparency and accountability. For example when a self-driving car crashes, who is responsible? Is it the car manufacturer, the insurer, the car owner, the other car driver, or the person that might have voluntarily jumped in front of the car to see how it reacts? This could happen in almost all verticals where AI is used. In any business and especially with the introduction of The General Data Protection Regulation (GDPR) in Europe, a customer could ask why they were rejected for a product or service. If you had used deep learning such as recurrent neural network (RNN) and had not done some analysis, it would be difficult to explain why they were rejected, as RNNs are a kind of a black box. In comparison, had you used a simpler decision trees it would be much easier to explain as the multi-level split are straightforward to understand.
This issue is also reflected in industry, for example Research from MIT, evaluated commercial facial recognition software, to see if it could correctly identify gender based on 1,200 photographs. It found that it was correct for 99.2% of the time for light skin males, but women with darker-skin tone were only correctly classified as female 65.3% of the time. Had this been used for more advanced identification scenarios, such a high error rate could have consequences on marketing but also on security, products, and services. Having more diverse teams to help balance requirements, ethics and belief is vital.
You could also miss out on investment or growth opportunities as you are not asking the right questions, especially if there is not much signal in the data or no data on a segment of the population. For example if you used beverage consumption data in the UK in the 1970’s to determine if you needed better coffee in continental Europe, you would probably find that people are mostly happy with tea and instant coffee. Companies like Costa Coffee were founded by Italians unhappy with the coffee in the UK compared to Italian standards in the 1970’s. If you took the same data in Italy in 1970’s, you would get a different view of the market. In 2016 Costa Coffee was in 3,401 locations worldwide with a revenue of £1.167b. It’s about asking the right question, overcoming preconception and being open for opportunities.
Q. How to ensure all traits are considered by a diverse team building a diverse AI future for the planet
I think having a diverse team is essential, I like to break it down to diversity of backgrounds, thought and personality. In a New York Times article called Tech’s Troubling New Trend: Diversity Is in Your Head, it talks about the importance of cognitive diversity. If you think about your own organisation or experience in the workplace, you will find that some teams need a quiet working environment for deep thinking, and some people will have headphones on, which can be contrasted to more creative and expressive teams. Going deeper on the topic, other studies found that individual experience and resilience are based on profession, geography, and organisation levels too.
Q. What do you recommend for women who want to work in AI now?
At the moment AI is still in its infancy and there are more opportunities and roles emerging. Use a support network such as Girl Guiding, do online courses (Coursera, Khan Academy etc.), and find your passion area.
Whether STEM or non-STEM/Humanities all can benefit and be involved to some degree.
For those in STEM whose passions, interest or experience are in software engineering, what I would recommend is thinking about the engineering roles in ML or data science that are very technical and need a background in computer science and software. The passion is needed as it’s very fast moving and some frameworks and knowledge can become outdated after even 6 months, even if the skills are transferable and code sometimes easier to write. So you need to be motivated and keen to self-learn, e.g. for deep learning it used to be all about Google’s TensorFlow still very popular, since 2017 there are easier alternative frameworks like MXNet supported by Amazon & Microsoft or Cafe2 and PyTorch supported by Facebook. Or you can use abstractions frameworks like Keras or Gluon.
For those in STEM who are not as experienced in software delivery, they can think about roles involving more analytical thinking, like an analyst or a data scientist. You may not be a developer, but I still recommend that you maximize your exposure to the software delivery process and languages such as SQL, Python or R. You can also look at openness, transparency, scrutinising and explaining the algorithms. For example if you are rejected by an insurance company you may ask why and what data they used?
For non STEM/humanities there are a lot of non-technical related roles such as data journalism, marketing, UX, and storytelling. There is more on the human side too if you think about the roles where AI is used and tested, in ethics and society. As AI gets built into robots, autonomous vehicles, and the workplace you need to think of outcomes and changes. The aim of AI is ultimately in using it in areas such as improving people’s lives, public protection, productivity, wellbeing, and healthcare.
Q. Are gender differences in Technology real?
I’ve worked with many talented men and women throughout my career, and I think anyone with the right passion and interest can work in Tech and AI, and it’s probably down to overcoming stereotypes and having the right opportunities.
Research by the Open University estimated that there are only 17% of women working in IT in UK, compared to 35% in India. Why? In India there are fewer stereotypes surrounding the type of people who undertake STEM careers, more collaboration between education providers and industry, and more encouragement from parents to pursue careers in tech. In the UK they say the issues are at the entry level, women usually do horizontal moves into IT when more senior. The study also says that in the UK stereotypes about the people who usually fill technology roles put girls off of choosing a technology career from a young age. Parents have a big role in supporting girls in their choices.
I would like to end with a study carried out by the University of Washington’s Institute for Learning and Brain Science (I-LABS), involved about 100 six-year-old girls and boys. The children were split into three groups: children who programmed animal-like robots, played a storytelling game, and with no activity. All three groups were then given a survey asking them of their opinion of technology and their beliefs about whether girls and boys had different abilities. They found equal interest for boys and girls.
“As a society, we have these built-in beliefs that are pushing boys toward certain activities more than girls. So our thought was, if you give equal experiences to boys and girls, what happens?” Master said. “We found that if you give them access to the same opportunities, then girls and boys have the same response — equal interest and confidence.”
“Stereotypes get built up in our heads from many different sources and experiences, but perhaps if we give girls more experience doing these kinds of activities, that will give them more resources to resist those stereotypes,” Dr Master said. “They might be able to say, ‘I can still be good at this and enjoy it, despite the cultural stereotypes.’”
My French international school’s moto was “enrich ourselves of our differences” and I think that new ideas and experiences from people in different backgrounds can be incorporated together to make a much richer product or service. Breaking down the stereotypes and giving everyone an opportunity to work in AI is the way to do this!
Connect with me on LinkedIn