Fears about how artificial intelligence (AI) and robots might transform our lives have been a staple of debate and science fiction for decades. But what is more interesting is AI’s broader interaction with current issues and how they are raising the stakes higher. AI and democracy, AI and sustainability, AI and nutrition, or AI and finance are some of the fields where machine learning is thriving and encouraging theoreticians and academics to reach out to other sectors and actors.
EPFL’s Applied Machine Learning Days (AMLD) aims to open up the debate on these future crossroads by creating a venue that is relevant to academics, practitioners and possibly the general public. It is a presentation of the state of the art as much as an opportunity to exchange, share and learn. The 2021 edition was launched online yesterday and will offer a series of events and discussions throughout the year, exploring why artificial intelligence in general, and machine learning in particular, is important, especially today in the context of the pandemic. As expressed by the co-organiser of the event, EPFL assistant professor and head of the Data Science Lab, Robert West, in his opening speech:
“Coronavirus has shown so many challenges and opportunities for applied machine learning, such as modelling the structure and the behaviour of the virus, modelling and predicting the viral spread, automated diagnosis of Covid, tracing in privacy preserving ways, developing vaccine logistics, measuring and tightening vaccine hesitancy, fake news detection. Machine learning can help fix the problems created by Covid-19. And of course, it only works if it is applied.”
AI and democracy. “This is democracy’s day,” US President Joe Biden declared in his inauguration speech last week, making the subject of AMLD’s first “track”, or session - AI and democracy - all the more timely.
With the revelations related to Cambridge Analytica and the massive use of fake news exacerbated by the Covid crisis, the potential threat of AI to democracy has become an important and debated issue. The co-organisers of the first track acknowledge the challenges but would also like to explore how machine learning and AI could help improve and enhance democracy. The examples are numerous: corruption and hate speech detection, campaigning and misinformation or preference and opinion learning.
But to understand and benefit from the potential of certain models, one must walk across sectors. Victor Kristof, one of the co-organiser of the event and PHD student at EPFL, explains the importance of his collaboration with University of Geneva political science lecturer and fellow co-organiser, Steven Eichenberger :
“I was developing these models that give us very high accuracy in predicting, for example, the success of votes. We also obtained some nice interpretation and visualisations of the results. But then we were a bit stuck on the “so what” question. What do we do with these predictions? What kind of questions are we able to answer with our predictive approach? That's where we need people like Steven, to help us integrate our methods in a broader context of political science. To really try to posit some hypotheses and to answer them from this consultative approach.”
This particular aspect is enhanced by the fact that approaches to the problematic differ in political science, says Eichenberger:
“Predictions are not really what we do in political science. We are more into explanation. But we are interested in voting predictions as they can be voting determinants. They can affect opinion formation, and the participation of voters. So in the beginning of our collaboration, I was amazed by the amount of data Victor had at his disposal and could see the potential. I believe that by digging in political science literature, machine learning models can improve.”
Moreover, AMLD is an opportunity to showcase the work that is being done and build bridges between communities.
Political ads on social networks. The concept is simple. A series of guest speakers are invited to share the state of their research and concrete solutions on the subject. Among them, CNRS research scientist, Oana Goga, presented an independent auditing system of facebook ads, exposing one of the challenges democracy is facing on social networks.
“We are witnessing a shift from using ads to promote products to potential interested people to target well chosen information for specific groups of people that are likely to accept it. And this ability to target information can bring new dangers for our democracy. The most famous example was Cambridge analytica.”
The events of 2016 made people realise that it is possible to weaponise technology, to engineer polarisation, to disengage and even to manipulate voters into choosing one candidate over the other. In response, both ad platforms and governments started to adopt countermeasures with a series of restrictions on so called political ads. The question is : how do we know an ad is political? The problem, Goga explains, is that advertisers need to declare on a voluntary basis if they are sending ads with political content. And there is no way to know what happens if advertisers do not self-label their ads as political. Hence, the development of an independent system to audit the ads on facebook. It is based on a net collector, which is a Chrome and Firefox extension that collects ads and microtargeting information, and the users themselves, who donate data about the ads received on Facebook.
The results of a first case study on the Brazilian elections of 2018 are promising.
“Two per cent of the ads people saw on Facebook were political, but were not declared as being political. This number must be compared with the fraction of official political ads in people's feeds, which is two to four per cent. This means that we have an equivalent number of declared and undeclared political ads.”
To move forward, Goga concludes, it is also up to the international community to help define what is and is not considered as a political ad.
Applied machine learning in reality. Times have changed and the reality of machine learning creates new opportunities in terms of technology development and concrete applications. West describes the evolution:
“I think the key development that we're seeing is that machine learning is becoming a technology. What I mean is that it actually works now. Earlier, it was research. But now, we're seeing this kind of productionisation and realisation of machine learning where you can actually have these tools that start to work. One example is chatbots. This is something that was completely utopian, even 10 years ago. What has also changed is the open source spirit. Companies like Google or Facebook make their models and code available which wasn’t the case before.”
The challenge is probably to bring understanding of this technology to companies, West continues, and fill the gap between the technology itself and the needs of the industry have in terms of solutions. But the AMLD also works the other way around. For academics it is essential to understand what works and doesn’t work in practice. And by presenting the mentioned topics this year, the ultimate goal is ambitious: machine learning for good.
The good means bringing the bright side of machine learning to its darker corners, which include the problem of data privacy. West continues:
“I like to call it sometimes learning without knowing. How can we learn useful things without ending up knowing everything about everyone. It is both in terms of what you can learn and about what you can infer out of the information that's already there, that can be sensitive. But also, the decisions that are then made based on the data can be harmful, because these algorithms are often quite dumb. They are trained on uncertain data that they see. And then they generalise that to unseen data. But most of the algorithms that we use don't have common sense and don't have any sort of ethical norms baked into them. So if the data is bad, you will get bad decisions.”
Learning without knowing starts with learning and bringing the questions to the table. By reaching out to the practitioners and the public, and crossing over fields of research, the AMLD create a space of learning and concrete solutions to take a step further in the direction of a safe and constructive association between machine learning and democracy.