EPFL Scientists Point to Perils of Unregulated AI

Through the years, artificial intelligence has progressed from being a futuristic application in secretive scientific projects to a technology that shapes almost all aspects of modern life. When it first arrived, there were intense discussions on the possibility of AI running amuck, relegating humans to becoming the slaves of AI-run machines. More than half a century of formal research since then, punctuated by major advancements, has established the benefits of AI and made it a defining technology for the future. However, as the applications of AI increase rapidly, the scientific community, as well as lawmakers, are concerned that unregulated AI can lead to misuse and abuse.

In this context, the European Commission took a clear stance by releasing a white paper in February 2020, which could lead to a regulatory framework for AI. That was followed up later in the year by the European Parliament, which adopted a set of proposals on how the EU can best regulate AI to boost innovation, ethical standards, and trust in technology. More recently, scientists convened at a virtual platform called the ‘Governance Of and By Digital Technology’ to deliberate on the principles required to regulate current and future digital technologies and prevent the harmful impact of decision-making algorithms. The event was hosted by EPFL’s International Risk Governance Center (IRGC) and the European Union’s Horizon 2020 TRIGGER Project.

Among the speakers at the conference were EcoCloud members Bryan Ford and James Larus. Professor Ford is Associate Professor at EPFL and head of the Decentralized and Distributed Systems Laboratory (DEDIS) in the School of Communication and Computer Sciences, while Professor Larus is Dean of the IC School and IRGC Academic Director.

Professor Ford called for a “cautious use” of powerful AI technologies but warned against their implementation in defining, implementing, or enforcing the public policy. Instead, policymaking should remain an exclusively human domain. Citing a real-world example, he said, “AI may have many justifiable uses in electric sensors to detect the presence of a car—how fast it is going or whether it stopped at an intersection—but I would claim AI does not belong anywhere near the policy decision of whether a car’s driver warrants suspicion and should be stopped by Highway Patrol.”

Professor Larus emphasized that AI applications have benefits as well as risks, and they must never be allowed to cross the “thin red line” between the two. Regulations are needed to maintain that balance between benefits and risks.

Other speakers at the conference included Stuart Russell, Professor of Computer Science at the University of California, Berkeley. He drew attention to the already evident risks from poorly designed AI systems, which include online misinformation, impersonation, and deception. Marie-Valentine Florin, Executive Director of the IRGC, reminded participants that artificial intelligence is not an end in itself but only a means to an end.

Read More

Rafael Pinot Receives EcoCloud Postdoctoral Research Award

Distributed computing is already playing a pivotal role in a wide range of applications including web-based communications and intelligent information systems. In keeping with this development, EcoCloud presented a research project award to Rafael Pinot for his ongoing work on controlling the spread of fake news or infectious diseases based on the theory of gossip protocols. Rafael just obtained a Ph.D. in Computer Science from Université Paris-Dauphine PSL and CEA LIST institute.

Rafael is working on the project with EPFL professors Rachid Guerraoui, Anne-Marie Kermarrec, and Carmela Troncoso. The research aims to control the dissemination of information in a peer-to-peer network, with a special focus on gossip protocols. Inspired by the literature on adversarial vulnerabilities of these processes, the research will “turn adversaries into algorithms”, essentially using system weaknesses to its advantage. It has two main objectives. First, it aims to design algorithmic solutions to stop or slow the spread of fake news. Second, it will develop efficient methods to detect the source of fake news. While carrying out the study, the researchers will initially investigate the setting where agents communicate using a gossip protocol without learning from each other (dissemination of information) and then adapt their findings to the setting where agents aim to collaboratively learn a machine learning model (peer-to-peer machine learning).

Rafael has a strong academic foundation in Applied Mathematics, having received a BSc and an MSc in the subject from Université de Sorbonne. His work mainly focuses on the security and privacy of machine learning applications. With a penchant for investigating complex data structures, he is particularly interested in the way randomization makes algorithms safer.

Rafael started his academic career in 2017 as a research intern at CEA LIST Institute, where he worked on graph theory and privacy in machine learning. In mid-2018, he received a JSPS fellowship to travel to Japan as visiting researcher at Kyoto University. He continued to pursue his interests in machine learning at Dauphine-PSL University where he also worked on game theory, adversarial machine learning, and trustworthy machine learning.

Read More

New Humanitarian Initiative Brings Digital Technology and Engineering Skills to the Table

Managing humanitarian aid is one of the most important problems in the modern world. It is also a very powerful and direct approach to make a lasting global impact. Keeping that in context, experts at ETH Zurich and EPFL–two of Switzerland’s leading federal institutes of technology–have joined hands with the International Committee of the Red Cross (ICRC) to launch the Engineering Humanitarian Aid initiative.

The partnership brings scientific and technological expertise to tackle three crucial dimensions of humanitarian aid: energy and the environment; data sciences and digital technologies; and personalized health and related technologies. Having worked together on numerous occasions in the past, the three partners will promote developmental cooperation in the field by harnessing digital technology and engineering skills.

Some projects are already underway. Among them is a project led by Carmela Troncoso, professor at EPFL’s Security and Privacy Engineering Laboratory and member of the EcoCloud faculty. The project addresses privacy issues related to biometric data, which seeks to identify the ICRC’s on-the-ground security and privacy requirements when processing beneficiaries’ biometric data, and to design a system that keeps these data confidential.

Another ongoing project uses artificial intelligence to map vulnerable populations. It seeks to compile information from satellite images and social media posts to estimate the size and density of these populations, as well as associated information such as settlement type and changes in population. The work is being conducted by Konrad Schindler and Devis Tuia (EPFL).

In the area of technology for personalized healthcare, a project led by ETH Zurich’s Stephan Wagner aims to augment the availability of medical equipment in conflict zones and reduce wastage.

The collaborators are likely to expand the program over the next year to initiate several joint projects to develop and distribute innovative products and processes. At EPFL, efforts under the initiative are being coordinated by the EssentialTech Center, while at ETH Zurich they are being managed by ETH for Development (ETH4D). The ETH Board has allocated CHF 5 million as seed funding for the projects over a two-year period (2021–22).

Read More

AI Helps Predict ‘Cytokine Storms’ in COVID-19 Patients

COVID-19 not only has a range of symptoms, but also varying impact on patients. In its early stages, doctors are unable to predict the progression of the disease because it depends on the interaction between the viral infection, the patient’s response, and the development of cardiovascular inflammation. To address that problem, a cross-disciplinary program spearheaded by EPFL seeks to empower both patients and doctors with an assistive, predictive, and personalized healthcare tool called Digipredict.

Digipredict is a pan-European research program involving universities, hospitals, and startups. It aims to develop a digital twin that can detect serious complications in COVID-19 patients by using breakthrough technology in the fields of artificial intelligence, smart patches, and organs-on-chips. Playing a key role in the project are EcoCloud faculty members David Atienza and Martin Jaggi, who also head the Embedded Systems Laboratory and the Machine Learning and Optimization Laboratory respectively. The two other EPFL laboratories involved in the project are the Nanoelectronic Devices Laboratory and the Laboratory of Movement Analysis and Measurement.

Digipredict detects the first signs of a ‘cytokine storm’ in high-risk COVID-19 patients, thus allowing doctors to act before it causes serious damage to the cardiovascular systems. Cytokines are proteins that play an important role in normal immune responses, but having a large amount of them released in the body all at once can be harmful.

Digipredict predicts disease progression by using a smart patch with integrated technology for collecting crucial medical data such as blood oxygen levels, breathing rate and body temperature. Nanosensors linked to an AI program track specific biomarkers that indicate any possibility of a cytokine storm. The data collected and analyzed by Digipredict allows doctors to make informative and timely decisions about the course of treatment.

Apart from EPFL, other institutions involved in the project are the University of Twente, ETH Zurich, IMEC in Belgium, Stichting Imec in the Netherlands, the Charité university hospital in Berlin, the University of Bern (through Inselspital), and three startups (Ascilion, EPOS-IASIS, and SCIPROM). The Center for Intelligent Systems (CIS) will be responsible for promoting Digipredict and disseminating its findings.


Read More

Pascal Frossard Named as Full Professor (STI)

The Board of the Swiss Federal Institutes of Technology has announced the appointment of Pascal Frossard as Full Professor of Electrical Engineering and Electronics in the School of Engineering (STI). Currently Associate Professor at EPFL, he joined the EcoCloud faculty in 2018 to help the research centre drives its cloud computing programs.

Before joining EPFL in 2003, Professor Frossard was a member of the research staff at the IBM TJ Watson Research Center at Yorktown Heights, NY, USA. His core research areas include interpretable machine learning, data science, graph signal processing, image representation and analysis, computer vision and immersive communication systems. His research contributions include the analysis of geometric properties of deep networks, deep nets robustness analysis, and representation learning for graph signals.

Professor Frossard has won international acclaim for expertise in signal and image processing, and its applications in intelligent systems and biomedicine. He is known for an innovative research approach that combines different disciplines in the natural and engineering sciences and facilitates essential academic and industrial partnerships. Over the years, he has won many distinctions. These include, inter alia, the Swiss National Science Foundation Professorship Award in 2003, IBM Faculty Award in 2005, IBM Exploratory Stream Analytics Innovation Award in 2008, Google’s 2017 Faculty Award, IEEE Transactions on Multimedia Best Paper Award in 2011, and IEEE Signal Processing Magazine Best Paper Award in 2016. He is a Fellow of IEEE. His work has been widely published in reputed journals. In his most recent publication,* the authors introduce a representation learning algorithm for graphs, which simultaneously learns a low-dimensional space and coordinates for the nodes in that space.

Professor Frossard’s appointment as Full Professor will undoubtedly strengthen EPFL’s key strategic research areas.

* Simou, Effrosyni & Thanou, Dorina & Frossard, Pascal. (2020). node2coords: Graph Representation Learning with Wasserstein Barycenters. IEEE Transactions on Signal and Information Processing over Networks. PP. 1-1. 10.1109/TSIPN.2020.3041940.


Read More

New “Sight and Sound” Algorithms to Identify COVID-19 Patterns

About a year ago, when the novel coronavirus broke out, medical science not only failed to arrest its spread but also to properly identify the developmental stages of the disease. Many casualties resulted because the progression of the disease was an enigma. In the later part of the year, there were nascent attempts to harness AI for COVID-19 diagnosis, treatment, and monitoring. A giant step in that direction has been taken recently by researchers at EPFL; they have developed algorithms that can practically see and hear COVID in a patient’s lungs.

The new deep learning algorithms DeepChest and DeepBreath have been developed by the team of Dr. Mary-Anne Hartley at EPFL’s intelligent Global Health group (iGH) based in the Machine Learning and Optimization Laboratory of Professor Martin Jaggi. DeepChest uses lung ultrasound images, while DeepBreath utilizes breath sounds from a digital stethoscope. The algorithms can accurately diagnose the novel coronavirus in patients and predict how much they will be affected by the virus strain.

Two nearby Swiss university hospitals are involved in developing the algorithms. At HUG, the Geneva University Hospitals, Professor Alain Gervaix has been collecting breath sounds since 2017 to develop an intelligent digital stethoscope called the “Pneumoscope” to diagnose pneumonia. The recordings have now helped develop the DeepBreath algorithm at EPFL. Initial trials show that DeepBreath can detect even asymptomatic COVID by identifying changes in lung tissue before the patient becomes aware of them.

The clinical aspects of the DeepChest project are being conducted at CHUV, Lausanne’s University Hospital. Thousands of lung ultrasound images are being collected from patients admitted to the Emergency Department with COVID-19 symptoms. Although the image sample collection process started last year, they have since focused on COVID-19.

The algorithms are available on the EPFL website, but they are very much a work in progress. Efforts are on to further refine and validate the algorithms by inviting coding skills from around the world, including a year-long hackathon called ‘CODED-19’ by the EPFL community. As Professor Jaggi explains, “This AI is helping us to better understand complex patterns in these fundamental clinical exams. So far, results are highly promising.” At a later stage, iGH plans to develop a mobile application for the deep learning algorithms and make them available far and wide.

While the current effort is to specifically meet the COVID-19 challenge, its preliminary success has amply demonstrated how large-scale AI research can be used to remove some of the roadblocks for medical science.


Read More

EPFL Scientists Spearhead Deep Learning Efforts to Clear Space Debris

[et_pb_section][et_pb_row][et_pb_column type=”4_4″][et_pb_text]

The space around the Earth is home to abandoned satellites, rocket boosters, and other bits of space debris that have accumulated over the years. It is estimated that there are tens of thousands of space debris in low Earth orbit (LEO). At least 34,000 orbiting pieces of junk are larger than 10 cm each, which make them a major threat because they are orbiting the Earth at very high speeds. There were two near collisions recently, which could have been disastrous: In September, the ISS was forced to conduct a maneuver to avert a possible collision with an unknown piece of space debris, while there were high chances of a Chinese ChangZheng-4c rocket crashing into an old Soviet Parus navigation satellite in October.

With numerous missions being conducted by several space organizations, space must be made a safer place for all future explorations. Toward that objective, technologies are being developed to capture and deorbit space debris. One such pioneering mission is ClearSpace, a spin-off stemmed from the EPFL Space Center (eSpace). The first mission of the project, ClearSpace-1 scheduled for 2025, is to recover the now obsolete Vespa Upper Part, a payload adapter orbiting 660 kilometers above the Earth that was once part of the European Space Agency’s Vega rocket.

At least three EPFL laboratories are working in tandem on the project: Computer Vision Laboratory led by Professor Pascal Fua, Realistic Graphics Lab led by Assistant Professor Wenzel Jakob, and Embedded Systems Lab spearheaded by Professor David Atienza (also a faculty member at EcoCloud).

ClearSpace-1 aims to extricate the debris from space by enabling the robotic arms of a captured rocket to approach the Vespa from the correct angle, grasp the object, and pull it back into the atmosphere. That calls for the development of deep learning algorithms to estimate the 6D pose (3 rotations and 3 translations) of the target based on video-sequences. That’s where the Computer Vision Laboratory comes in with inputs from Mathieu Salzmann, the lead scientist working on the project.

However, the scientists must allow for physical changes in the Vespa over the seven years of its space wanderings. To estimate its current appearance for Professor Salzmann, the Realistic Graphics Lab is generating a database of synthetic images of the target object, which includes a detailed 3D model of the Vespa upper stage.

The nature of the project requires the scientists to ensure completion of the mission in space with limited computing power onboard. The Embedded Systems Lab is leading that segment of the work of transferring the deep learning algorithms to a dedicated hardware platform. As Professor Atienza explains, “Since motion in space is well behaved, the pose estimation algorithms can fill the gaps between recognitions spaced one second apart, alleviating the computational pressure.” However, he adds, “…the algorithms are so complex that their implementation requires squeezing out all the performance from the platform resources.”

Therefore, the key to the success of the mission is to design algorithms that are 100% reliable in space. And therein lies the challenge, as well as the inherent attraction, of the entire project for the scientists at EPFL.




Read More

Anne-Marie Kermarrec Joins EcoCloud

EcoCloud is happy to announce that Anne-Marie Kermarrec, Professor in the School of Computer and Communication Sciences at EPFL, has now come aboard the EcoCloud faculty.

Anne-Marie Kermarrec is a widely acclaimed computer scientist with deep academic and business experience. After completing her PhD thesis at Rennes (France) in 1996, she worked with the computer science group at Vrije Universiteit in Amsterdam for a short while before taking up the position of Assistant Professor at the University of Rennes. Between 2000-2004, Anne-Marie was part of Microsoft Research in Cambridge. For the following decade, she worked as a Research Director at INRIA in Rennes. In 2015, she founded Mediego, a startup company for online personalized predictive services that directly leveraged her research work.

Anne-Marie has won several honors and awards. She received an ERC grant for GOSSPLE (2008-2013) and an ERC Proof of Concept for the new recommender AllYours (2013). She received the Michel Monpetit Award (2011) and the Innovation Award (2017) from the French Academy of Science. She has been elected to the European Academy in 2013 and named ACM Fellow in 2016. She has also been on the program committee of many workshops and conferences and served as Chair of the ACM Software System Award for a number of years.

Since January 2020, Anne-Marie has made a significant academic contribution as Professor at EPFL. Her research investigates large-scale distributed systems and more precisely P2P systems, epidemic algorithms, distributed infrastructures for machine learning, and privacy-aware personalization systems.

While welcoming Anne-Marie Kermarrec to the EcoCloud faculty, we are sure that her international stature as a hardened researcher will amplify the center’s activities and its leading academic role in the global arena.

Read More

Hédi Fendri Wins OMEGA Prize Student Award

Each year, a student of the Microengineering Section of EPFL is awarded the prestigious OMEGA Student Award for contributing to scientific and technological advances in the disciplines of Microengineering, Micro- and Nanotechnologies, and Chronometry. This year, the Board of the Foundation of the OMEGA Prize decided to honor Mohammed Hédi Fendri, a master student at the Institute of Microengineering in EPFL’s School of Engineering.

Hédi’s Master’s thesis “ML-Based Side-Channel Analysis and Disassembly of Hardware Root of Trust” is based on his findings in the industry itself. His research crossed path with the Kudelski Group, a Swiss company that has established itself as a global leader in digital security and convergent media solutions for the delivery of digital and interactive content. At Kudelski, Hédi was supervised by Jérôme Perrine and Marco Macchetti, Director (Engineering) and Principal Engineer, respectively. Hédi’s academic supervisor was Mirjana Stojilovic, a faculty member at EcoCloud.

In his thesis, Hédi devised a methodology for assessing the side-channel vulnerability of a device at design time and demonstrated its efficiency on a RISC-V CPU design, simulating the execution of the AES-128 encryption. The thesis also presents a novel side-channel analysis disassembler based on dictionary learning and sparse coding. Unlike previous works, which targeted 8-bit microcontrollers running on frequencies up to 16 MHz only, the proposed ML-based method can successfully recognize all instructions of a 32-bit RISC-V CPU running on frequencies as high as 100 MHz.

Hédi’s findings are both important and timely. With the rise of the IoT, smart home appliances, embedded medical devices, and other numerous IoT devices have become an attractive target for adversaries. Hédi’s findings enable significant reduction in cost and time necessary to accurately assess the side-channel vulnerability of an IoT device well before its deployment.

Laureates of the OMEGA Student Awards receive 1000 Swiss francs, usually in addition to an OMEGA watch. The award ceremony is held concurrently with the graduation ceremony for Master’s degrees in Microengineering at EPFL. The Board determines the number of awards every year; Hédi is the sole winner in 2020. We congratulate Hédi on receiving this award and wish him every success in his future endeavors.

Read More

Award Ceremony: ICCAD 2020 Ten Year Retrospective Most Influential Paper Award

The ICCAD Executive Committee has recognized “3D-ICE: Fast Compact Transient Thermal Modeling for 3D ICs with Inter-Tier Liquid Cooling” as “the most influential on research and industrial practice in computer-aided design of integrated circuits over the ten years since its original appearance at ICCAD.” The authors of the paper are Arvind Sridhar, Alessandro Vincenzi, Martino Ruggiero, David Atienza (all from the Embedded Systems Laboratory – ESL at EPFL), and Thomas Brunschwiler (IBM Zurich Research Laboratory).



Read More