Synced canvases the Swiss artificial intelligence landscape: from software research in Lugano to hardware development in Zurich, with supporting vertical markets spanning the country’s west side, backed by EU investments and reaching consumer markets in Germany and around the world.
To begin surveying the artificial intelligence landscape in Switzerland, we should first consider the country’s unique economic and political structure. The Swiss economy is built on the self-governance of its 26 cantons, governing 260 municipalities. Cantonal governments have their own constitutions and policies separate from the federal establishment. Each is also in charge of dealing with cross-border business, natural resources, higher education, cantonal taxes, and citizenship — all factors that can make the AI industry thrive or shrivel. The Swiss business culture reflects a way of thinking deeply-rooted in organic and bottom-up politics, is closely-knit, and highly functional.
Switzerland is a cultural mosaic, with a 63% German-speaking population residing in the north and northeast, 22.7% French-speakers in the southwest, and 8.4% Italian-speakers in the south. It borders Germany, France, Italy, Austria and Liechtenstein. The multiculturalism, historical neutrality and political stability of Switzerland have helped anchor its prosperous financial industry and open the door to trade opportunities and cross-border talent exchanges.
Swiss science dominates the fields of physics and mathematics, as backed by stable endowments from the Swiss National Science Foundation and talents from influential EU facilities like CERN. The country has the highest number of Nobel laureates per capita (ignoring tiny statistical outliers St. Lucia and Iceland), and maintains its lead in applied research as the world’s most innovative country, ranked by the World Intellectual Property Organization. Swiss universities are liberal and international, which attracts German researchers who come from a much more ethnically uniform environment.
The country’s AI research has a solid base in the Italian-speaking city of Lugano, extending westward to the French-speaking cities of Martigny, Geneva, Lausanne, Delémont, and stretching around the country’s perimeter to the German-speaking cities Basel and Zurich. Together these cities describe a “C-shaped” arch around the country, with research labs concentrated along the borders with Germany and France. In this survey, we’ll walk through top-notch Swiss AI facilities: starting with deep learning and neural network research at Dalle Molle Institute for Artificial Intelligence Research (IDSIA) in Lugano, to interdisciplinary research at École Polytechnique Fédérale de Lausanne (EPFL) and University of Basel, and ending with robotics innovations at ETH in Zurich and University of Zurich. Along the way, we’ll introduce the reader to several important researchers and their projects, experiences, and perspectives.
IDSIA in Lugano and Deep Recurrent Neural Networks
The Dalle Molle Institute for Artificial Intelligence (IDSIA) is located in the industrial outskirts of Lugano, one of the most scenic lakeside cities in the Italian-speaking canton Ticino. In 1988, Swiss philanthropist Dalle Molle set up a foundation to fund three research institutions dedicated to AI research: IDSIA in Lugano, IDIAP in Martigny, and another in Geneva.
Today, IDSIA is affiliated with two high-level institutes, Università della Svizzera Italiana and University of Lugano, housing around 50 researchers, including a dozen PhD students.
The lab has made research breakthroughs in many areas, including deep learning, adaptive robots, and ant colony optimization to name a few. Dr. Jürgen Schmidhuber is renowned for the invention and subsequent development of long-short-term memory (LSTM) in the 1990s, a recurrent network algorithm that helps machines learn many things unlearned by feedforward networks. A demonstration of LSTM can be found on everyone’s mobile phone. “Since mid 2015, Google’s speech recognition is based on LSTM, with forget gates for recurrent units trained by ‘connectionist temporal classification (CTC)’. This approach dramatically improved Google Voice […] by almost 50% — now available to billions of smartphone users,” explains Schmidhuber. Since 2009, his team has received over USD 12 million in academic funding, including a recent Europe Research Council (ERC) grant.
Schmidhuber wants us to understand that despite the present hype ignited by Alpha-Go’s victory over Go master Ke Jie, feedforward networks were created as early as 1975 by Ukrainian Soviet scientist Alexey G. Ivakhnenko, who was able to train an eight-layer neural network with much-limited computation power. The future possibilities of this method fascinated Schmidhuber, who was doing his PhD in Munich, Germany in the 1980s. “I studied maths and computer science. For the cover of my 1987 diploma thesis, I drew a robot that bootstraps itself in seemingly impossible fashion. The thesis was very ambitious and described the first concrete research on a self-rewriting “meta-program” which not only learns to improve its performance in some limited domain, but also learns to improve the learning algorithm itself, and the way it meta-learns the way it learns etc., improving itself endlessly, except for the limits of computability identified by Kurt Gödel in 1931.”
Schmidhuber’s aim has been to build general purpose “embodied intelligence,” which is not just the application of AI to gaming and simulations, but to physical robots at large. His lab also applies algorithms to robots, including the adorable humanoid “iCub” robot programmed by IDSIA to pick up cans; drones that travel in forests using image recognition; and swarm robots that can jointly accomplish tasks.
Some of the research at IDSIA is being continued at spinoff companies like DeepMind, whose co-founder Shane Legg was a PhD student in Dr. Schmidhuber’s group, along with other early core-members. The startup was sold to Google for USD 600 million in 2014.
Interdisciplinary research at EPFL and Basel
Moving west to Canton Vaud we find the École Polytechnique Fédérale de Lausanne (EPFL). Lausanne has a very solid research foundation in STEM education, with intensive focus on robotics, and two prominent labs including an Intelligence System Lab and Biorobotics Lab. It is also known for housing the Human Brain Project that started in 2013, a ten-year EU flagship research project with an operating budget of EUR 1.019 billion.
Boi Faltings joined EPFL in 1987 as professor of Artificial Intelligence, and heads a lab there with six PhD students. Says Faltings, “The artificial intelligence research in Switzerland is very design and engineering centric. When I started to work on software in the 1990s, I worked on the diagnostic systems for jet engines and also optimized airport scheduling systems.” However, some of the systems did not perform as expected. As Dr. Faltings recalls, working with enterprises is sometimes problematic: “Enterprise need short-term solutions, new research on the other hand is very risky for companies, and the projects didn’t end up being implemented.”
Faltings watched the Swiss artificial intelligence research community shrivel during the AI winter, while serving as president of the Swiss Group for Artificial Intelligence and Cognitive Science (SGAICO), an organization active in the 1980s organizing conferences for domestic researchers. Following revived interest in deep learning in recent years, Faltings saw “the field boom from 1000 to 5000 researchers within the span of three years. Most people don’t know that much about AI, they will leave when things stop working.”
In recent years, Faltings has spent a lot of time working on self-organizing agents and on discerning truthful datasets using machine learning. His only concern is for students leaving academia: “Nowadays everyone works for just a few big companies. The data is also concentrated in the hands of a few big companies, but should belong to the public.”
Upon Dr. Faltings’ introduction, we were connected with Dr. Malte Helmert in Basel, who works alongside a group of 30 researchers at the University of Basel in the German-speaking canton of Basel-Stadt. The current focus of this group is on domain-independent automated planning and heuristic state-space search, where they are “interested in algorithms that can solve the general planning problem without being tailored by a human expert and without requiring a learning phase.” In the last five years, Helmert’s group of 10 researchers received CHF 3.5 million from an Economic Research Council (ERC) grant, an EU collaboration project, and the Swiss National Science Foundation.
Other notable institutions that conduct AI basic or applied research along the C-arch are the IDIAP lab in Martigny, University of Geneva, University of Nauchatel, University of Applied Sciences Western Switzerland (HESOS), and Lucerne University of Applied Sciences and Arts.
Robotics in ETH and University of Zurich
In northern Switzerland, ETH Zurich is one of the best universities in the world conducting robotics research, with eight labs and about 120 PhD students. The flying robots of Raffaello D’andrea of the Dynamics Systems and Control Lab are sensational to watch, while the work of Bradley Nelson on microrobotics and soft robots at the Institute of Robotics and Intelligent Systems has earned a broad international reputation. Around 30 students conduct research at Dr. Roland Siegwart’s Autonomous Systems Lab.
During our visit, Siegwart’s PhD students were working on drones that fly over crops, collecting biomass development information across the field. The project runs in conjunction with Bosch, the cross-sector Swiss giant in appliances. The drones relay real-time data for farmers’ decision-making, cutting time and labour costs. The robotics lab also collaborates on projects with Google, Microsoft, and Siemens.
The lab’s recent advancements are in Simultaneous Localization and Mapping (SLAM), a technology that helps a robot navigate its environment through either laser, sonar, or visual data input. A recent breakthrough is a method called Open Keyframe-based Visual-Intertial SLAM (OKVIS), which helps the robot sync bits of visual information with sudden movements across small time-intervals. OKVIS was first published in The International Journal of Robotics Research in November, 2014.
While there are public concerns with the advancement of industrial robots, Siegwart tells us that he is less worried about the impact of automation regarding job losses, as “society will adapt and there are environments where people should not have worked in to begin with.” There are, however, other concerns: “The physical world has many complications. Imagine if your smartphone crashes, you will buy another smart phone. But if an autonomous car breaks down people will be killed. So far we are doing research in confined environments, but as soon as drones are flying over people, they should have failing modes and be extremely aware of people.”
The difference between robotic research in Switzerland and United States as opposed to Japan, is that the former “have a much more functional view of machines,” explains Siegwart, befitting the sober and rational Swiss temperament. Yet there are exceptions. For example, in collaboration with Disney’s research lab, ETH created BeachBot to draw sand cartoon patterns on beaches to entertain passers-by.
ETH Zurich collaborates with many institutions, including Germany’s Max Plank Institute and Stanford University in the US. It also conducts joint research programs with the University of Zurich, with which it shares a campus. Dr. Rolf Pfeifer, professor emeritus at the University of Zurich and the head of its AI research lab before it shut down in 2014, tells us he is among the very few in Switzerland with an interest in humanoid robots.
Pfeifer is one of the earliest Swiss AI researchers. After experimenting with Freudian dream simulations and a series of unfruitful trails with the expert system, he made a mid-career shift from algorithm research to robotics in the 90s — which are “very different domains” from software research — coming to the conclusion that artificial intelligence has to be a stand-alone organism. “The concept I’ve been researching is ’embodiment’, which studies the roles of the body in the development of intelligence. The robots’ acoustic, visual, and tactile senses help them learn from the real-environment, generating patterns of sensory information. Much of the current research in algorithms has zero interaction with the real world.”
Pfeifer is content with the support for his work, “I have always had very comfortable funding, 80% came from the European Union. There’s a program called Future and Emerging Technologies (FET) that has a really good funding scheme. The European Research Council takes a bottom-up approach, listens to researchers’ recommendations and then comes up with budgets. The Swiss National Science Foundation, on the other hand, provides medium-sized funding, but not as much.”
Pfeifer initiated the National Centre of Competence in Research Robotics (NCCR) along with Dr. Dario Floreano from EPFL. “These are really people doing the most interesting work, biologically inspired robots, wearable robots, neuroprosthetics, soft robots, etc.” The centre integrates ETH Zurich, University of Zurich, EPFL, and IDSIA with industry partners, providing innovation funds to startups in multiple fields of robotics. He hopes to see more interdisciplinary work in the near future, combining neural science, cognitive science, material science and robotics.
Switzerland’s Competitive Advantage: Industrial Robots and Industry 4.0
Industry 4.0 is a concept coined by the German government to expedite the current trend of automation and data exchange in manufacturing technologies. In 2015, there were 248,000 industrial robots sold, with 68,000 purchased by China alone. As predicted by the Boston Consulting Group, robots can cut costs at least 20%. By 2025, machines will take over 23% of the manufacturing workload currently performed by humans. For mature industries like automobiles and electronic appliances, reducing manufacturing costs is a strategy that ensures business success.
Switzerland’s competitive advantage in industrial robotics lies in its complete industry value chain. According to Economic Performance of Swiss Regions by the University of Fribourg’s Centre for Competitiveness, there are significant concentrations of ICT and robotics upstream and downstream businesses along the south and northwest border. This includes industries like metal, watches, material science, microtechnology and high-tech equipment, alongside academic institutions conducting robotics research. Particularly, the country’s traditional strength in the watch industry helped lay a strong foundation in precision engineering, while intensive academic research on robotics technology creates high entry barriers for global competitors.
Swiss industrial robotics companies also have business models different from their international competitors. They have the capability to do research, find supplies, finish manufacturing, assemble in-house and provide customer service to global clients. This differs from Japanese manufacturers for example, who outsource assembly and customer service to downstream companies through public bidding, American companies meanwhile mainly do assembly and non-standardized customization. This explains why the Swiss industrial robotics company ABB is capable of locking in profits and scale internationally.
With headquarters in Zurich, ABB is by far the most successful international player in industrial robots, with 132,000 employees worldwide and revenue of USD 34 billion. A few years ago, the company started providing digital services and software on top of selling hardware products. In 2016, service and software contributed 18% of ABB’s revenue.
The other industrial robot conglomerate with a strong Swiss presence is German company Kuka AG, recently acquired by China’s Midea Group. Kuka works with the Swisslog group based in Buch, a city near Zurich, with two business units: Healthcare and Warehouse & Distribution Solutions (WDS). Swiss companies like Menzi Muck, Schindler the elevator giant, and German auto companies such as Volkswagen are also direct beneficiaries of the progress of industrial robotics. Smaller players include robotics startup companies — mainly spinoffs from ETH Zurich and EPFL that produce service robots, flying drones and so on.
Swiss AI: From Startup to Ecosystem
What does it really feel like for AI startups in Switzerland? We brought the question to Lugano and spoke to the co-founder of NNAISENSE, Jan Koutnik, former post-doc from IDSIA with a research focus on artificial neural networks, recurrent neural networks, and deep reinforcement learning. His former advisor, Dr. Jürgen Schimidhuber is a co-founder of the company.
NAISSENSE was founded in 2015. Today, much of the company’s environment is reminiscent of the IDSIA lab, where a group of smart guys strive to find business solutions with AI. There are currently 16 employees, but NNAISENSE may expand to 24 people soon, opening a new branch office in Texas, USA. When I visited, Koutnik was working on an Audi AG project using reinforcement learning algorithms to reduce autonomous driving vehicles’ parking noise made by turning tires.
Jan Koutnik holding an an Audi Q2 model at NNAISENSE’s office in Lugano.
“For Audi, they figure that if they outsource the task to us, we can do it more efficiently,” says Koutnik. NNAISENSE’s first subsidiary, Quantenstein, was created with German fund Actis, building financial portfolios with adjustable parameters to maximize profit. The company doesn’t deal with direct products, rather they provide solutions to big cross-sector corporations throughout Europe.
Koutnik, originally from the Czech Republic, has a quibble with Switzerland’s current immigration policy, which could be keeping talent out. His former lab mate Shane Legg, originally from New Zealand, co-founded DeepMind in London, partially because he had to leave Switzerland after graduation.
Aerial view of Lugano, the Italian speaking part of Switzerland. Image retrieved via Wikimedia.
In February 2014, an initiative on “mass immigration” was passed in Switzerland to impose immigration quotas by 2017. This hinders the effect of EU’s 2002 agreement on the Free Movement of Persons within member countries. On one hand, an open immigration policy brings in highly-skilled migrant workers, especially from Germany. But with five land borders the country is susceptible to large-scale involuntary migration, according to a recent World Economic Forum report on national risks.
Despite the immigration issues, the overall artificial intelligence ecosystem in Switzerland is robust. In 2016, Google opened Europe’s largest artificial intelligence lab at Google Zurich. IBM has had centres in Zurich for many years, producing four Nobel Laureates. Thomson Reuters recently opened an incubator in Zurich, “giving startups working in areas such as big data, analytics, AI and distributed ledgers access to its framed data and content sets.”
Ultimately, all commercial ecosystems are supported by their monetary flow. Until very recently, the Swiss had been conservative with their investment decisions, with over half of startup growth financing coming from abroad, most notably German venture capital investors. According to the Swiss Venture Capital Report, “the median of all financing rounds remains relatively low at CHF 2.5 million” (roughly USD 2.6 million). The best financed companies were in biotech and medicine tech, which “raised about CHF 577 million in 2016 compared to CHF 484 million in 2015.” They were followed by ICT and fintech companies which attracted CHF 271 million, around 30% of the Swiss-wide volume.
The numbers are not so impressive when compared to the USA, China, and Canada’s investment in AI technology. However, it is important to understand Switzerland’s strong national appeal to cross-border commerce. The country draws talent from the European Union, houses the United Nations Office in Geneva, and has always taken the lead in basic and applied academic research. Following Brexit, Switzerland is one of the few European countries capable of becoming the next financial centre. UBS and Credit Suisse are giants holding approximately half of all domestic banking assets. The country’s reputation also draws international attention, as Swiss quality and innovation is not a marketing technique, but a mutually agreed upon fact. Internationally entrusted biopharmaceutical companies such as Hoffmann-La Roche and Norvartis are conducting vast bioinformatics research. The Swiss-German partnership in upstream research and downstream industrialization is also an important global partnership.
As Dr. Schmidhuber puts it, “Switzerland always had a history of immigrants doing unexpected things at unexpected places. Earlier, German Albert Einstein revolutionized physics while working as a patent clerk at Bern. and British physicist Tim Berners Lee invented the World Wide Web at CERN.”
Regarding the Swiss quality of life, Koutnik tells us that he enjoys his picturesque lakeside city very much and does not plan to leave anytime soon. The balance of work and lifestyle is great for his family. “It’s really nice here, I wouldn’t want to go anywhere else.”
Many thanks to Boi Faltings, Jürgen Schmidhuber, Roland Siegwart, Rolf Pfeifer, Malte Helmert, Jan Koutnik, Raghav Khanna, Zhen Gao for contributing their experiences and technical expertise to this article.
This is the second article within featured series AI and Globalization, click to read Building AI Superclusters in Canada.
Author: Meghan Han | Editor: Michael Sarazen | Producer: Chain Zhang