The success of predictions depends on the availability of sufficient and high-quality data. Better predictions will affect all industries. In Switzerland, we want to take advantage of good predictions, but are not willing to pay for infrastructure. Today, special applications of artificial intelligence already have superhuman capabilities. For example, they can detect errors in a production line. The challenge is to adapt these applications for the masses. The decryption of the black box AI is also central: we have to know why certain decisions are better than others. The prediction of black swans is in the distant future. We can solve more and more complex problems, but there will be even more complex problems.
The central question is: what does my idea reduce the cost of? The success of Artificial Intelligence can be explained in a simple sentence: AI reduces the cost of predictions. When predictions become cheap, we define problems as prediction problems that were not previously defined as such. So the if-then problem of autonomous driving became a prediction problem: what would a human do behind the steering wheel? Today, predictions help HR departments, the development of translation software, In giving medical diagnoses, in the granting of loans, in fraud detection. The circle of success is working for big companies like Google: more data means better predictions. The improved offering attracts more users, who in turn deliver more data. The ingredients are there, that predictions become the next "Sputnik moment". The investment sums are very high, China is investing heavily. Now everyone is faced with a decision: wait, or join in.
Predicting a heart attack by means of data evaluation is no longer science fiction. If patients have the symptoms of kidney disease, it is too late and they have to undergo dialysis. In the future, we will already know in advance who will soon have kidney disease. From reactive medicine to proactive, preventive medicine. Artificial Intelligence that analyses data sees what people cannot see. Preventive medicine also makes sense economically, it is cheaper. The problem is: we are dinosaurs, we are afraid of change. But we have to adapt, otherwise we won't have a chance at some point. Doctors will not be replaced by algorithms, but their job profile will change. Doctors will still be important for human interaction. As a result, their education has to change as well.
Swiss Re is an aggregator of all the risks that arise in the world. Thanks to predictions using artificial intelligence, risks such as health risks or natural catastrophes can be insured much more individually. The risk landscape of tomorrow is shaped by macro trends. The Swiss Re Sonar Report 2019 has identified risks and trends in medicine such as genetic therapies, water shortages, the urban-rural divide, the logistics centre of the future and climate change. Today it is possible to simulate a person's state of health in 20 years and thus define their insurance risks. Algorithms are only human: they can be manipulated and can also be wrong.
Netflix is one of the most successful companies in the world and one of the most data-driven. This is no coincidence. Netflix had nearly 150 million subscribers in 190 countries at the beginning of 2019. They are responsible for 15% of the Internet's total downstream traffic. In 2018, Netflix spent $12 billion on content, and by 2019 it is expected to invest $15 billion. Prediction models tell the team how many users will watch which content, then they invest. Using big data evaluations, Netflix also produces its own films. Netflix continuously tests its website and conducts A/B tests. Each section of the site is personalised for each individual user. But AI only helps to a certain extent. Employees should ultimately make the decisions.
Today we all live in a computerised world where everything can be calculated and analyzed. The goal of control through predictions is human behavior. It is becoming more difficult to predict the future because the effects of phenomena such as climate change are difficult to predict. Prediction reduces the ability to make one's own decisions. Search engines and social networks are prediction machines. However, they predict on the basis of our preferences and drive us to ever more extreme positions. Fear and anger are the result. Also, because we have to trust and follow the decisions of the machines, but do not understand them. Computers do not give us the answers for a better world. It is up to us to make the right decisions.
Together with the "Stiftung Risikodialog" and the GDI, the participants played a card game that enables a discussion of the opportunities and risks of prediction algorithms. The game revealed the possibilities of communication and discussion about the implementation of future technologies.
At the Power of Predictions conference, one of the three afternoon sessions was entitled "Predictions in Financial Services and Beyond. Hosted by Swiss Re Institute, this was a fascinating look across the globe into the potential, the challenges and the limitations of predictions. The three experts on hand were Evangelos Avramakis (Head Digital Ecosystems R&D, Swiss Re Institute), Sandra Andraszewicz (Researcher and Project Leader, Behavioral Finance Team, ETH Zurich) and Christian Klose (Senior Analytics Professional, Swiss Re) and they offered participants three inspired rapid-fire presentations.
Evangelos Avramakis shared insights into just how quickly technology has evolved. He looked at examples of prediction models from healthcare to mobility and from the retail sector to natural catastrophes Despite the massive value that can be generated for humanity with the help of AI-driven technology, he also pointed out dangers of increasing information asymmetry becoming a growing economic and social concern as well as political powers undermining democratic principles.
Sandra Andraszewicz talked about the limits of predictability of human financial decisions. For 300 years scientists have aimed to develop accurate models that would help describe, explain and predict human financial decisions. But while the age of digitalization and big data has led to a breakthrough in collecting and analyzing behavioral data, Sandra Adraszewicz showed that algorithms are often blind to human intentions and unrecorded variables.
In the third presentation, Christian Klose explored how Deep Learning techniques can help predict wildfires. The societal and economic risks of wildfires have become more concerning in recent years, especially in densely populated areas and areas with industrial operations. Christian Klose showed how Deep Learning techniques allow for the prediction of wildfire occurrence, severity and patterns months in advance – thus helping decision makers take preventive action.
This breakout session provided attendants the opportunity to gain insights into state-of-the-art research on AI-powered predictions technologies for enterprises which are explainable, secure and fair as well as robust and scalable. Today, there are already very powerful and pervasive applications of prediction technologies. However, they focus on narrow use cases in a specific field for one specialized task. These range from equipment failure predictions to demand predictions in the fashion industry, to predicting the onset of epileptic seizures. In addition, while in some contexts, time is not a critical aspect, in others, for example in medical and emergency response applications or autonomous driving, real-time prediction, alerting, and action is essential. Abdel Labbi, IBM Distinguished Engineer and Researcher, highlighted some of the latest research in technological and algorithmic advancements, including so-called in-memory computing as well as novel ultra-fast automated machine learning and training methods. These help to speed-up prediction machines and make them more scalable and accurate – in this way also allowing for more complex models and applications.
Anika Schumann, Manager AI for Industries and Services, and Dorothea Wiesmann, Department Head for AI Research at IBM’s Zurich Lab, focused in their talks on explainability, fairness and trust, which are all crucial aspects for the acceptance and use of predictions technologies in industry and society – eventually determining their overall impact. Up until now, prediction systems have been developed as black-boxes, without explaining why the system suggests something to a specific user. With recent developments in regulations and privacy requirements around data and AI, users now have the Right to Know and the Right to Contest any data-based decision that they may receive. For new use cases in healthcare, education, energy, finance, recruiting, mobility and in many more fields, explanations are important to make predictions actionable and to build trust. With technologies such as the open-source IBM AI 360 Fairness toolkit, it is for example already possible to detect, quantify and mitigate biases in data sets used for training AI models. The toolkit is also helping AI developers to see inside their creations via a set of dashboards and dig into why they make decisions – and there are more efforts at IBM Research aimed at providing real-time insights into algorithmic decision making.
The idea of using animals for predictions is not new. The new thing about Wikelski's projects is to link the information animals give him to a "Internet of Animals". Animals can predict zoonotic diseases and their spread, weather and climate, crop failures and natural disasters. Wikelski equips the animals with transmitters to measure their behaviour and nutrition. From elephants to goats and birds of prey to bees and butterflies: Wikelski uses all these animals for his predictions, in real time all over the world. The Icarus project has set up a receiving station on ISS for the signals from the transmitters. Natural intelligence is much more interesting and efficient than artificial intelligence. And it has proven itself over millions of years.
Markets become data-rich, supply and demand become a better match. Markets become what they always wanted to be: efficient means of coordination. They become more competitive against companies. Superstar companies are markets in the guise of companies that set the framework for transactions. Data is the new money. Data takes over the information function of the price. Data-rich markets have a big problem that we can already see today: monopolisation. The reason for this is the feedback effect: systems that learn from data become better and better the more feedback data they generate. We need a new antitrust law for the digital age and data-rich markets. This leads to a data tax so that everyone can benefit. Data capitalism: The next stage of capitalism has begun. Our generation's task is to shape it in such a way that the market creates added value for all.
Computer-aided predictions make sense when it comes to making life normal again, such as in fighting diseases or protecting the climate; even consumption is about normalisation. However, the fall of the Berlin Wall, 9/11, Fukushima have shown: Everything really important are black swans, the absolutely unpredictable. The aberrations are the interesting thing. The question is: How can artificial intelligence be used to augment natural intelligence? In the field of living things, there are functional mechanisms that are more effective than machines. Robustness helps, not predictions. We must learn to love the unexpected.