Machine learning is essential for a resilient future
The future will be ever more intensely data-driven. Raw data have very little value by themselves – they have to be curated and transformed into insights that lead to value-creating decisions. Useful transformed data requires both data engineering and some kind of machine intelligence, such as machine learning. Valuable transformed data will (and in many cases already do) underlie and connect everything and everyone. Unfortunately, in the financial sector, data have still not been harnessed in a transformative way. By working together and striving for trusted collaborations across all layers of technology, we will be able to transform our industry - and Swiss Re Institute's research partnerships play a key role in this endeavor.
This is just the beginning
At the recent Applied Machine Learning Days in Lausanne, Switzerland, I highlighted that 85% of all AI projects fail. Based on my experience, that number is even higher. In terms of machine learning, we still find ourselves at the very beginning of the digital revolution. This is a point in time that is as exciting as it is daunting – success depends on a few core factors.
From failures to success
At this early stage there are still basic foundational challenges. How we define 'artificial' and 'intelligence' will influence research, development and deployment of machine intelligence." The data science community needs to a) define and agree on a global machine intelligence taxonomy, and b) translate those efforts for decision-makers in business and government. Once there is a common understanding, there is common ground to move forward in a transformative way.
AI projects will see a far greater chance at successful implementation, if they give a particular focus to four areas, namely Augmented Intelligence, Intelligent Automation, Assessed Intelligence and Adaptive Intelligence. Augmented Intelligence focuses on augmenting and thus improving the productivity of humans. Intelligent Automation is essentially about building systems that integrate humans and machines in productive ways (instead of just replacing humans entirely with machines). Assessed Intelligence is a key area that is all about making models robust by evaluating them rigorously and continuously. Finally, Adaptive Intelligence means creating more resilient systems that can adapt to changing circumstances by shifting to a causal-inference paradigm. (view keynote presentation)
Humans in the loop
Key to the aforementioned AI areas is human intelligence. What I'm most interested in is figuring out how we can keep humans in the loop, making the best of what both machines and human beings have to offer. Successful implementation of AI-driven systems within corporations very much depends on this. On one hand, full automation invariably clashes with company culture and leads to mistrust, barriers and blockers. On the other hand, human beings deliver on both experience and imagination in ways machines cannot.
During the AMLD Conference, Swiss Re Institute ran a dedicated track on 'AI and Resilience'. One of the keynote speakers, Nicole Hu, Co-Founder and CTO of One Concern, explained that "modeling is best combined with human insight and decision-making." In addition, when it comes to machine intelligence, human involvement is a key risk mitigator. An AI might pursue objectives that are not in the best interest to humans, delineating from an original purpose. Humans have the ability to think in alternatives beyond the usual. Counterfactual imagination may be the key to mitigating machine intelligence’s risks.
The algorithmic risk challenge
Algorithms already play key roles in everything from mortgage decisions to insurance product pricing, and from mobility systems to healthcare. Most people are unaware that they are an integral part of our everyday lives. As the world becomes ever more interconnected, the algorithmic use of data needs to be monitored: algorithmic risk will become, along with cyber risk, one of the biggest threats we face in society.
Problems, both in terms of fragility and vulnerability, will increase as systems become more complex. Today, data scientists freely avail themselves of model components from GitHub, Stack Overflow and similar platforms. Have these components, as they are incorporated in interconnected models around the world, been properly evaluated? The answer is, all-too-often, no. For the most part, algorithmic risk is still too rarely addressed. There are insurance implications and I would like to see many more of us work on this key emerging risk.
Our data-driven future must be built on collaboration. Swiss Re Institute engages in dozens of research partnerships with institutions and businesses around the globe. Among these is the long-standing engagement with EPFL, the Swiss Federal Institute of Technology and the EPFL's C4DT (Center for Data Trust), a partnership that brings together twelve partners and thirty-four laboratories – all focused on delivering trust-building technologies.
The Swiss Re Institute's 'AI & Resilience' track at the AMLD conference also featured Nicole Rieke (Senior Deep Learning Solution Architect at NVIDIA). She focused her keynote on Federated Learning, a machine learning technique that trains an algorithm across decentralized systems – while keeping the data protected. Rieke's healthcare-centered work uses federated learning to collaboratively learn across institutions, while keeping patient data safe. The example perfectly illustrates both the power of collaboration and the possibilities of working together in trusted, decentralized networks. The future doesn't lie with traditional centralized networks – the future, and thus also trust, is distributed across decentralized networks.
While an ever more data-driven future is assured, the future itself remains elusive. However, we're human and thus we have what it takes to intelligently integrate machines into our lives. We have intelligence, we have imagination – and we have conscience. We are perfectly capable of learning from our failures, focusing on the right challenges, managing and monitoring new risks, and collaborating toward shared goals. Again, we have what it takes.
The tasks at hand, and ahead, are all-encompassing, global challenges. Swiss Re has a crystal-clear purpose: We make the world more resilient. Swiss Re Institute supports this mission as a conduit for sharing useful knowledge and fostering better decision making. Artificial intelligence, developed and deployed in the best interests of humanity, will help us deliver on our purpose. A data-driven future will allow us to predict and prevent events. Instead of just absorbing shocks and helping people, businesses and governments recover, we will be able to make the world safer by offering resilience-as-a-service.
We have every reason to be hopeful – if we work together in tackling the above, our data-driven future will be one that is socially, economically and environmentally far more resilient.