How to take a human centred approach in building Machine Learning Solutions
--
Human Centred Design Principles lie at the heart of building Ethical and Responsible AI and technology solutions
Ethical AI means Human Centred Design
Human-centred design (HCD) is a design philosophy that enables tech teams build technology that has as its focus the abilities, needs and interests of people who are affected by this technology. This is directly related to the movement to build Responsible AI. The end goal of how to make technology development more ethical lies in discerning how to place at the heart of technological development, the interests and well being of people and society as a whole.
As everything that I have covered in this post is directly applicable to the question of Responsible and Inclusive AI, it’s useful to draw the link between the two. AI Ethics and Fairness is a great deal more than mitigating bias in data and algorithms: we need to consider thoroughly how to centre people, local communities and the humanity involved in the generation to build a just digital future. Human Centred Design (HCD) as a philosophical stance gives us a framework to allow us to do exactly this.
What is Human Centred Design?
The Human Centred Design or HCD approach means respecting the agency of all the people involved in the design and build process: from the conception of the technology solution, to its deployment and continued monitoring. It means respecting equally the various skill-sets — beyond just coding and mathematics, that it takes to build a truly ethical solution to a real world societal issue. ‘Ethical’ in this regard means a solution that respects the humanity, and latent agency of the creators of this solution, as well as those this solution is meant to help. I propose, this should be the first and the only guiding principle in designing a solution — and not whether it would be the fastest, most profitable . It should also take priority over the interest in a solution that showcases the most innovative technology. In my opinion, we cannot hope to create technology that is in alignment with general human welfare without these principles — even if it means sacrificing on speed of innovation and production.
The HCD method therefore requires a fundamental rethink about design processes and expected outcomes.
As HCD tech teams confront a problem, they would start by thinking about whether a particular problem needs, can, or should have a technical solution at all — and whether a manual, often community-based approach might not be better after all. Community-based approaches would take a participatory approach to understanding what “a solution” might look like to a particular community. This means empowering community actors to design and propagate the technology generation process. Agency over how and for what purpose data is used should also be completely democratic. And finally accountability around how well the solution is working should be decentralised. Access to the information of how these solutions are being deployed or how they perform after deployment should be transparent and easily accessible.
So tell me how you really feel…
Let’s walk through a practical example of this. Imagine an ad company that has a database of millions of videos of people responding to a particular campaign. Until now this ad company had hired a person to review these videos to figure out what kind of emotion the respondents are displaying in reference to the advert. To cut costs, they would like to automate the process by training a computer vision algorithm to recognise respondents’ emotions — a process called “sentiment analysis”. This situation raises ethical concerns on several levels.
- Lack of consent / Mis-use / Data colonisation : The first, most pressing one is that this company has access to video data exhibiting personal, human emotion. They are now monetising the data without the consent of the participants in order to sell products.
- Inaccurate: The second issue is with the automation process. Emotions and how we express them are largely a very specific, cultural and lived experience-dependent process. Even humans, who each individually possess all of this lived experience, get it wrong. How many social situations have we all been in where we just didn’t “read the room right”? Therefore, training a computer algorithm to do this would not only be technically difficult — but also technically unsound. The realm of possible emotions and associated ‘feeling labels’ that this algorithm would have to be trained on must be infinite . A person very rarely feels just “happy” or just “sad”, but more often, a mixture of conscious and subconscious emotions.
- Discriminatory: Moreover, if an ad company takes this automated approach, chances are that the algorithm trying to recognise emotions on faces will only have been trained on faces belonging to a very specific subset of the population. Given current socio-economic histories and forces, these faces in the training data would very likely be those that have been socialised in a Western context. Their emotions and means of expression would be relevant only to a Western context, but the trained algorithm would be deployed on people from all cultures and backgrounds. The problem of biassed datasets is not restricted to emotion recognition. Take for example the benchmark face datasets IJB-A and Adience critiqued in Dr. Buolamwini and Dr. Gebru’s paper “Gender Shades: Intersectional Accuracy Disparities in Commercial Gender Classification”. Though these were often used to commercially train algorithms to recognize faces, the datasets had an overwhelming majority (79.6% and 86.2% respectively) of lighter skinned faces. This led the facial recognitions algorithms reporting very low accuracy on darker skinned female/femme-presenting faces.
How would a HCD Scientist approach this problem?
A HCD approach to this would first evaluate the necessity for automating this process of sentiment analysis. In this case, the HCD technologists would conclude that automating the process at this time posed too many risks to inclusive social and human interests. A sophisticated technical solution might be tempting from the perspective of speed and “efficiency” — where we might be socialised to believe that a technical solution must be the most accurate. In the case of sentiment analysis of the respondents to a marketing campaign, a technical solution would be inaccurate, thereby inefficient, and downright dangerous.
As a general rule, tasks which are repetitive and pattern-recognition based are well suited to being automated through a Machine Learning based solution. Even this has caveats, as Machine Learning algorithms are in general quite bad at understanding data that they haven’t exactly seen before. An example of this is when a classic neural network trained to label pictures of cats and dogs, often mislabels pictures of cats and dogs when they are really blurry or the picture is taken from a different angle to the training images. They are easily recognisable to the human user but not to the algorithm. How to fix this is an open and exciting area of research called Anomaly detection or understanding out of distribution patterns. Watch this space for more on this — and possible ways of mitigating bias in algorithms using this new research.
Is it time to throw in the towel with tech or…?
Absolutely not!
I am a scientist and technologist and I remain passionate about developing technical solutions, and sharing skills between as diverse a group of people as possible. Instead, I am advocating for a people-first, instead of a technology-first approach. I am asking for us to be intentional about how and why we develop new tech, and for whom we want this tech to be deployed. And finally, I am hoping that we all, in our very different and unique ways start collaborating on creating a just, inclusive and sustainable future.
