UNREAL: 'These faces were generated by AI'
In this Thought Piece, Dr Ayu Saraswati, NSSN Machine Learning Engineer, contemplates the future of inclusion and diversity in the age of Artificial Intelligence.
Artificial Intelligence (AI) has entered the mainstream: it can generate a person that does not exists or an image just based on a text description. It convincingly de-aged Mark Hamill in a Star Wars spin off show without 3D image rendering. It can recognise objects, understand speed, and avoid objects travelling simultaneously making self-driving cars operationalised in the United States. It can [and does] predict your taste in movies or music and shapes your buying behaviour by tracing your online footprint. It can predict protein fold and potentially be used to design new medications. AI, has a transformative power to improve our lives, or does it?
Image: the faces included in the image above were produced by a Random Face Generator. Source: https://this-person-does-not-exist.com/en
Most AI algorithms in operation are based on a machine learning approach where the AI is shown many examples of the data it needs, to be able to predict or recognise different patterns. The machine is trained to figure out the features or patterns associated with the archival data. If you show it enough representative data, the algorithm generates a model that can recognise or predict the outcome of events it has never seen before.
Data is the key for these AI algorithms to work and to work WELL. So, where do the data come from? It is mostly historical data and human-generated data and if history has shown us anything, it is that people have inherent biases and flaws. Joy Buolamwini has reported in a congressional hearing that existing face recognition models are heavily biased towards Caucasian male faces and perform worst at recognising female persons of colour. Speech recognition models tend to favour male voices. If you have a hard time with your voice-activated assistant, try lowering your register.
Currently, the available data has big gaps in diversity and inclusion. In the book, “Invisible Women: Exposing Data Bias in a World Designed for Men”, Caroline Criado-Perez details how data collection for women is lacking and sometimes non-existent. One example is in-car safety data. Females are often more likely to be injured or die in car accidents because the safety features in cars are tested mostly on male crash test dummies. Anatomically correct or even proportionally correct female test dummies did not exist until very recently. Smaller male crash test dummies were used in place of women. Even when female crash test dummies are used, they are rarely in the driver’s seat. If this biased data is used to train AI to predict crash injuries, it will be very likely to be inaccurate for women.
Even when gender and race are not part of the data, inherent unconscious bias still affect the outcome. A triage AI was trained to prioritise patients based on how much treatment they needed. However, the underlying data had inherent biases on the race and economic line. It ended up prioritising rich Caucasian patients because the poorer patients often could not afford treatment and were underrepresented in the data. The result was devastating with many patients not getting the treatments they needed.
Does it mean AI is not living up to its potential in the real world environment? It does not have to be. The answer to a more equitable and fair AI is inclusive and diverse data sets. A local Swedish government agency took a second look at their snow clearing strategy. Previously, the municipality would clear the main roads first and then walkways last. However, the data showed women are more likely to walk, cycle, and take public transport to a greater extent than men. This resulted in pedestrian injuries that increased costs to 4 times the cost of the plough. They then reverse the order of snow clearing and there were far fewer injuries with no additional snow ploughing cost. AI algorithms perform much better when trained with better data. Programs like NSW AI assurance framework are aiming to develop and operate AI technology fairly and responsibly.
The inclusivity and diversity in the AI development teams who are behind the technology development is as equally important. Diversity in the teams leads to not only reduced inherent biases but also innovation and excellence. The technology industry is notorious for having a predominantly Caucasian male workforce. There are many initiatives to increase women's participation and other minority talents in STEM but we’re still light years away from achieving diversity. Women only account for 20-30% of the STEM workforce in Australia. Aboriginal and Torres Strait Islander people are extremely under-represented in STEM with only 1 in 200 having university-level STEM qualifications.
AI is a powerful tool for us to wield. It depends on us to train it responsibly and correctly for it to be able to live up to its full potential.
Dr Ayu Saraswati joins the NSW Smart Sensing Network as one of our engineers. She has been working in machine learning and data analytics for five years. Her thesis was focused on machine learning application to network intrusion detection, specifically to introduce transparency and visualisation to improve and gain actionable insights about the environment. She has also worked at the University of Wollongong on neural network-based compression for high sparsity and dimensionality data, as well as reconciliation of multiple data sources to populate enterprise architecture model using frequent sequence mining.
Learn more about Ayu here.