BIFOLD Colloquium 2022/02/14

Home >

BIFOLD Colloquium 2022/02/14

Apache Flink – From Academia into the Apache Software Foundation and Beyond

Speaker: Dr. Fabian Hüske

Venue: Virtual event

Time and Date: 4:00 pm, 14 February

Registration: If you are interested in participating, please contact: coordination@bifold.berlin

Abstract:

Apache Flink is a project with a very active, supportive, and continuously growing community. For several years in a row, Flink has been among the top ten projects of the Apache Software Foundation with the most traffic on user and development activity. Looking back, Flink started as a research prototype developed by three PhD students at TU Berlin in 2009. In 2014, the developers donated the code base to the ASF and joined the newly founded Apache Flink incubator project. Within three years, Flink grew into a healthy project and gained a lot of momentum. Now, almost 8 years later, the community is still growing and actively developing Flink. Moreover, it has established itself in the industry as a standard tool for scalable stream and batch processing.

In this presentation, Fabian Hüske will discuss Flink’s journey from an academic research project to one of the most active projects of the Apache Software Foundation. He will talk about the academic roots of the project, how the original developers got introduced to the ASF, Flink’s incubation phase, and how its user community and developer base evolved after it graduated and became an ASF top-level project. The talk will focus on the decisions, efforts, and circumstances that helped to grow a vital and welcoming open source community

Speaker:
(Copyright: private)

Fabian Hüske is a software engineer working on streaming things at Snowflake. He is a PMC member of Apache Flink and one of the three original authors of the Stratosphere research system, from which Apache Flink was forked in 2014. Fabian is a co-founder of data Artisans (now Ververica), a Berlin-based startup devoted to fostering Flink. He holds a PhD in computer science from TU Berlin and is the author of “Stream Processing with Apache Flink”.

Tracking Spooky Action at a Distance

Home >

Tracking Spooky Action at a Distance

Tracking Spooky Action at a Distance

The use of AI in classical sciences such as chemistry, physics, or mathematics remains largely uncharted territory. Researchers from the Berlin Institute for the Foundation of Learning and Data (BIFOLD) at TU Berlin and Google Research have successfully developed an algorithm to precisely and efficiently predict the potential energy state of individual molecules using quantum mechanical data. Their findings, which offer entirely new opportunities for material scientists, have now been published in the paper “SpookyNet: Learning Force Fields with Electronic Degrees of Freedom and Nonlocal Effects” in Nature Communications. 

Naostrukturen von Molekülen
Being able to predict and model the individual steps of a chemical reaction at the molecular or even atomic level is a long-held dream of many material scientists.
(Copyright: istock.com/peterscheiber.media)

“Quantum mechanics, among other things, examines the chemical and physical properties of a molecule based on the spatial arrangement of its atoms. Chemical reactions occur based on how several molecules interact with each other and are a multidimensional process,” explains BIFOLD Co-Director Prof. Dr. Klaus-Robert Müller. Being able to predict and model the individual steps of a chemical reaction at the molecular or even atomic level is a long-held dream of many material scientists.

Every individual atom in focus

The potential energy surface, which refers to the dependence of a molecule’s energy on the arrangement of its atomic nuclei, plays a key role in chemical reactivity. Knowledge of the exact potential energy surface of a molecule allows researchers to simulate the movement of individual atoms, such as during a chemical reaction. As a result, they gain a better understanding of the atoms’ dynamic, quantum mechanical properties and can precisely predict reaction processes and outcomes. “Imagine the potential energy surface as a landscape with mountains and valleys. Like a marble rolling over a miniature version of this landscape, the movement of atoms is determined by the peaks and valleys of the potential energy surface: this is called molecular dynamics,” explains Dr. Oliver Unke, researcher at Google Research in Berlin.

Unlike many other fields of application of machine learning, where there is a nearly limitless supply of data for AI, generally only very few quantum mechanical reference data are available to predict potential energy surfaces, data which are only obtained through tremendous computing power. “On the one hand, exact mathematical modelling of molecular dynamic properties can save the need for expensive and time-consuming lab experiments. On the other hand, however, it requires disproportionately high computing power. We hope that our novel deep learning algorithm – a so-called transformer model which takes a molecule’s charge and spin into consideration – will lead to new findings in chemistry, biology, and material science while requiring significantly less computing power,” says Klaus-Robert Müller.

the movement of atoms is determined by the peaks and valleys of The potential energy surface
Simplified two-dimensional depiction of the potential energy surface of the atoms C2H4O. The actual potential energy surface is 15-dimensional. Areas with low potential energy are depicted in blue; those with high potential energy in red. The black line depicts the reaction from ethanal (left) to ethenol (right).
(Copyright: Oliver Unke)

In order to achieve particularly high data efficiency, the researchers’ new deep learning model combines AI with known laws of physics. This allows certain aspects of the potential energy surface to be precisely described with simple physical formulas. Consequently, the new method learns only those parts of the potential energy surface for which no simple mathematical description is available, saving computing power. “This is extremely practical. AI only needs to learn what we ourselves do not yet know from physics,” explains Müller.

Spatial separation of cause and effect

Another special feature is that the algorithm can also describe nonlocal interactions. “Nonlocality” in this context means that a change to one atom, at a particular geometric position of the molecule, can affect atoms at a spatially separated geometric molecular position. Due to the spatial separation of cause and effect – something Albert Einstein referred to as “spooky action at a distance” – such properties of quantum systems are particularly hard for AI to learn. The researchers solved this issue using a transformer, a method originally developed for machine processing of language and texts or images. “The meaning of a word or sentence in a text frequently depends on the context. Relevant context-information may be located in a completely different section of the text. In a sense, language is also nonlocal,” explains Müller. With the help of such a transformer, the scientists can also differentiate between different electronic states of a molecule such as spin and charge. “This is relevant, for example, for physical processes in solar cells, in which a molecule absorbs light and is thereby placed in a different electronic state,” explains Oliver Unke.

The publication in detail:

Oliver T. Unke, Stefan Chmiela, Michael Gastegger, Kristof T. Schütt, Huziel E. Sauceda, Klaus-Robert Müller: SpookyNet: Learning force fields with electronic degrees of freedom and nonlocal effects. Nat. Commun. 12(7273) (2021)

Abstract

Machine-learned force fields combine the accuracy of ab initio methods with the efficiency of conventional force fields. However, current machine-learned force fields typically ignore electronic degrees of freedom, such as the total charge or spin state, and assume chemical locality, which is problematic when molecules have inconsistent electronic states, or when nonlocal effects play a significant role. This work introduces SpookyNet, a deep neural network for constructing machine-learned force fields with explicit treatment of electronic degrees of freedom and nonlocality, modeled via self-attention in a transformer architecture. Chemically meaningful inductive biases and analytical corrections built into the network architecture allow it to properly model physical limits. SpookyNet improves upon the current state-of-the-art (or achieves similar performance) on popular quantum chemistry data sets. Notably, it is able to generalize across chemical and conformational space and can leverage the learned chemical insights, e.g. by predicting unknown spin states, thus helping to close a further important remaining gap for today’s machine learning models in quantum chemistry.

Benchmarking neural network explanations

Home >

Benchmarking neural network explanations

Benchmarking neural network explanations

Neural networks have found their way into many every day applications. During the past years they reached excellent performances on various largescale prediction tasks, ranging from computer vision, language processing or medical diagnosis. Even if in recent years AI research developed various techniques that uncover the decision-making process and detect so called “Clever Hans” predictors – there exists no ground truth-based evaluation framework for such explanation methods. BIFOLD researcher Dr. Wojciech Samek and his colleagues now established an Open Source ground truth framework, that provides a selective, controlled and realistic testbed for the evaluation of neural network explanations. The work will be published in Information Fusion.

Imagine a doctor giving a life changing diagnosis based on the prediction of a neural network or an employer denying an application relying on the prediction of a neural network. From an end-user perspective it is not only desirable but necessary and often even legally required to accompany a model’s decision with an explanation in order to trace it back to the decisive parts of the input. These models must be based on genuinely solving a given problem, and not on exploiting spurious correlations found in the data. Explainable AI (XAI) research has recently developed various techniques to uncover the decision making process of the model. Beyond that, XAI bears also the potential to help improve model performance and efficiency, or to enable new data-driven scientific discoveries.

In the vision domain, the explanation of AI models can take the form of a heatmap, where each pixel in an input image gets assigned a relevance value or score, indicating its relative contribution to the final decision. So far XAI methods along with their heatmaps were mainly validated qualitatively via human-based assessment, or evaluated through auxiliary proxy tasks such as pixel perturbation, weak object localization or randomization tests. Due to the lack of an objective and commonly accepted quality measure for heatmaps, it was debatable which XAI method performs best and whether explanations can be trusted at all. 

Benchmarking neural network explanations
Top left is a picture and a question concerning the picture. The neural network analyzes both and answers the question. The answer is then explained with LRP (layerwise relevance propagation) and produces a heatmap, showing which pixels contributed significantly to the answer. The researchers compare this heatmap with a ground truth (GT) mask, which knows where to look in order to answer the question correctly. Comparing the heatmap and the GT mask, the researchers can objectively calculate how much of the relevance in the heatmap falls on the “object of interest” in the mask. In this case it is 98%.
(Copyright: Wojciech Samek)

Wojciech Samek and his colleagues propose instead to evaluate explanations directly against ground truth object coordinates using a restricted setup of synthetic, albeit realistically rendered, images of 3D geometric shapes. “To the best of our knowledge, this is the first ground truth based and realistic testbed for the evaluation of neural network explanations proposed in the literature,” says Wojciech Samek. The evaluation is based on “Visual Question Answering”.  “We train a complex neural network that is able to answer questions about images. We give the model a generated image with different objects (3D shapes). Since we have full control over the generation of the image – which object is where – as well as the generation of the question, we know exactly what the model has to look at to answer the question. We know what is relevant. For example: The gray background contains no information at all and should not be marked as relevant by a reliable, explanatory method. With this method we can not only measure exactly how good an XAI method is. We can also compare them for the first time really objectively while performing a difficult task.” The CLEVR-XAI dataset, that the researchers provide open source, consists of 10.000 images and 140.000 questions, divided into simple and complex questions. Simple questions pertain to one object in the image, while complex questions can pertain to multiple. “Another key contribution of our paper is the systematic comparison of the eleven most widely used XAI techniques,” explains Wojciech Samek. “We were extremely happy to find that the LRP (Layer-wise Relevance Propagation) method, that was established in a BIFOLD project between Klaus-Robert Müller and myself, was asserted and could establish itself among the best performing methods.”

The CLEVR-XAI dataset and the benchmarking code can be found on Github.

The publication in detail:

Leila Arras, Ahmed Osman, Wojciech Samek: CLEVR-XAI: A benchmark dataset for the ground truth evaluation of neural network explanations. Inf. Fusion 81: 14-40 (2022)

Abstract

The rise of deep learning in today’s applications entailed an increasing need in explaining the model’s decisions beyond prediction performances in order to foster trust and accountability. Recently, the field of explainable AI (XAI) has developed methods that provide such explanations for already trained neural networks. In computer vision tasks such explanations, termed heatmaps, visualize the contributions of individual pixels to the prediction. So far XAI methods along with their heatmaps were mainly validated qualitatively via human-based assessment, or evaluated through auxiliary proxy tasks such as pixel perturbation, weak object localization or randomization tests. Due to the lack of an objective and commonly accepted quality measure for heatmaps, it was debatable which XAI method performs best and whether explanations can be trusted at all. In the present work, we tackle the problem by proposing a ground truth based evaluation framework for XAI methods based on the CLEVR visual question answering task. Our framework provides a (1) selective, (2) controlled and (3) realistic testbed for the evaluation of neural network explanations. We compare ten different explanation methods, resulting in new insights about the quality and properties of XAI methods, sometimes contradicting with conclusions from previous comparative studies. 

Science and Startups launches AI Initiative

Home >

Science and Startups launches AI Initiative

Science and Startups launches AI Initiative

Science & Startups is the association of the four startup services of Freie Universität Berlin, Humboldt-Universität zu Berlin, Technische Universität Berlin and Charité – Universitätsmedizin Berlin. It provides a gateway and access to the joint programmes and resources of these universities, to successfully start and develop a company. Since 2021, Science & Startups has been specifically strengthening research transfer in the field of Artificial Intelligence (AI). Now they officially launched their new focus programme: K.I.E.Z. (Künstliche Intelligenz Entrepreneurship Zentrum). With an initial 6.85 million euros provided by the BMWI and co-financed by the state of Berlin, a unique ecosystem is being created. K.I.E.Z. is an initiative dedicated to facilitate entrepreneurs in AI with scientific expertise as well as access to capital, industry partners and hiring talent. K.I.E.Z. will be carried out in close cooperation with the Berlin Institute for the Foundations of Learning and Data (BIFOLD).

Representing the four startup services that launched K.I.E.Z.: Volker Hofmann (Humboldt-Innovation GmbH), Karin Kricheldorff (Centre for Entrepreneurship at TU Berlin, Marcus Luther (BIH Innovation), Dr. Tina Klüwer (Director AI, K.I.E.Z.), Steffen Terberl (Profound Innovation).
(Copyright: K.I.E.Z./Tanja Schnitzler)

In his statement BIFOLD Co-Director Prof. Dr. Volker Markl emphasized that this initiative is “yet another important building block on the way to establish Berlin in the Champions League of AI locations”. The programme will focus on an AI-oriented expansion of the entire innovation chain: from the identification of startup potential in research to the targeted acceleration of the feasibility phase to an accelerator programme at the new AI Campus Berlin. The integration of a strong industry network in all innovation phases will be a core element as well as the establishment of an AI Academy for startups and stakeholders. Thus, AI-based startups will be identified, optimally supported and established on the market.

On December 2nd, Dr. Tina Klüwer, director of the AI program K.I.E.Z., former founder and CEO of parlamind GmbH as well as board member of the German Federal Association for AI, and Dr. Susanne Perner, technology scout (with AI focus) at TU Berlin’s internal startup service – the Centre for Entrepreneurship, will give an introduction to the new initiative.

Date and time: December 2nd, 2021, 12:30 pm – 13:30 pm

Location: virtual

Register: here

More information:

Artificial Intelligence Entrepreneurship Center, K.I.E.Z.

United against Cyberattacks

Home >

United against Cyberattacks

United against Cyberattacks

BIFOLD Researchers Prof. Dr. Georgios Smaragdakis and Prof. Dr. Anja Feldmann, together with colleagues from Deutsche Commercial Internet Exchange (DE-CIX) and Brandenburg University of Technology, show that the exchange of information about ongoing cyberattacks has the potential to detect and mitigate substantially more attacks and protect critical parts of the Internet infrastructure. Their paper “United We Stand: Collaborative Detection and Mitigation of Amplification DDoS Attacks at Scale” was accepted at the ACM Conference on Computer and Communications Security (ACM CCS) 2021.

DDOS ATTACK " and Alert icon on display of computer for management server in data server room
There is currently no effective solution to mitigate DDoS attacks.
(Copyright: istock.com/Anucha Cheechang)

Our commercial and social activity is increasingly moving online. The ongoing pandemic is a reminder of how dependent we are on the smooth operation of online services such as e-banking, e-commerce, video streaming, video conferencing, and e-mail. Unfortunately, attacks on such critical services are at an all-time high. The so-called Distributed Denial of Service (DDoS) attack is a popular type of cyberattack that aims to make servers unavailable. 

“At least once a week, we read in the news about a DDoS attack that shuts down online services of enterprises or the public sector. We are very frustrated that 20 years after the first DDoS cyberattacks, current solutions are not effective enough to mitigate such attacks.”, says BIFOLD Fellow Georgios Smaragdakis, Professor of Cybersecurity at TU Delft. The researchers collaborated with eleven network infrastructure operators, Internet Exchange Points (IXPs), in central and south Europe and the USA, which interconnect more than 2.100 networks.

They developed inference techniques to detect DDoS cyberattacks and applied them on massive network data collected at the distributed vantage points over six months. In total, they detected and analyzed more than 120.000 cyberattacks. Their results show that between 500 and 1.500 DDoS cyberattacks with traffic of more than one Gigabit per second are observable every day. The total attack traffic reaches up to four Terabytes per day (see Figure). Unfortunately, 80 percent of these attacks are not detectable with current DDoS detection techniques applied locally at each site!  

In total the researchers detected and analyzed more than 120.000 cyberattacks in six months.
(Copyright: Prof. Dr. Georgios Smaragdakis)

Closer investigation shows that DDoS cyberattacks have become more sophisticated. Today, attackers target many applications (ports) in parallel and can generate higher attack traffic levels at a lower cost. Moreover, attackers can use compromised machines around the globe to target a victim service. Thus, current DDoS detection techniques that analyze data at one location fail to detect them as they lack the global view of the ongoing attack. Even if local techniques detect some of the attacks, the detection is late and yields a higher cost of mitigation. “Without the large-scale network data collected at different vantage points, it would not have been possible to understand why DDoS detection mechanisms are not any more effective. The same data also shows that a given attack is visible at multiple locations. Thus, if the involved network infrastructures had informed each other about ongoing DDoS cyberattacks, they could have jointly defended against them”, explains BIFOLD Fellow Prof. Georgios Smaragdakis.

Motivated by their findings, the researchers developed and tested a DDoS Information Exchange Point (DXP) where network infrastructure providers exchange data about ongoing attacks, including attack traffic level, sources, and destinations of attacks in a trusted environment. The evaluation of the proposed DXP shows that smaller network infrastructure providers that may lack resources and expertise to defend against cyberattacks are benefitted the most. With DXP, mitigation of cyberattacks comes at a lower cost. Indeed, DDoS attacks are detected up to ten minutes earlier, and the attack traffic is dropped closer to the source of the attack. The DXP is now under testing and is expected to be fully operational in the following months and help to protect thousands of networks and services around the globe against DDoS attacks. 

The publication in detail:

Daniel Wagner, Daniel Kopp, Matthias Wichtlhuber, Christoph Dietzel, Oliver Hohlfeld, Georgios Smaragdakis, Anja Feldmann: United We Stand: Collaborative Detection and Mitigation of Amplification DDoS Attacks at Scale. CCS 2021: 970-987

Abstract

Amplification Distributed Denial of Service (DDoS) attacks’ traffic and harm are at an all-time high. To defend against such attacks, distributed attack mitigation platforms, such as traffic scrubbing centers that operate in peering locations, e.g., Internet Exchange Points (IXP), have been deployed in the Internet over the years. These attack mitigation platforms apply sophisticated techniques to detect attacks and drop attack traffic locally, thus, act as sensors of attacks. However, it has not yet been systematically evaluated and reported to what extent coordination of these views by different platforms can lead to more effective mitigation of amplification DDoS attacks. In this paper, we ask the question: “Is it possible to mitigate more amplification attacks and drop more attack traffic when distributed attack mitigation platforms collaborate?”

To answer this question, we collaborate with eleven IXPs that operate in three different regions. These IXPs have more than 2,120 network members that exchange traffic at the rate of more than 11 Terabits per second. We collect network data over six months and analyze more than 120k amplification DDoS attacks. To our surprise, more than 80% of the amplification DDoS are not detected locally, although the majority of the attacks are visible by at least three IXPs. A closer investigation points to the shortcomings, such as the multi-protocol profile of modern amplification attacks, the duration of the attacks, and the difficulty of setting appropriate local attack traffic thresholds that will trigger mitigation. To overcome these limitations, we design and evaluate a collaborative architecture that allows participant mitigation platforms to exchange information about ongoing amplification attacks. Our evaluation shows that it is possible to collaboratively detect and mitigate the majority of attacks with limited exchange of information and drop as much as 90% more attack traffic locally.

Scheduling computing tasks can reduce emission

Home >

Scheduling computing tasks can reduce emission

Scheduling computing tasks can reduce emission

Flexibility is key to the successful integration of variable renewable energy sources into the power grid.  Matching electricity demand to current supply, often referred to as demand-side management, is already practiced by major energy consumers such as warehouse cooling systems and steel production factories. Data centers worldwide also consume large amounts of energy, currently well over 200 terawatt-hours per year. This demand corresponds to approximately 40 percent of the total German energy consumption and is expected to rise further. To reduce the carbon footprint of cloud computing, researchers from the Berlin Institute for the Foundation of Learning and Data (BIFOLD) investigated the potential of shifting delay-tolerant compute workloads, such as batch processing and machine learning jobs, to times where energy can be expected to be green. Their publication “Let’s Wait Awhile: How Temporal Workload Shifting Can Reduce Carbon Emissions in the Cloud,” was now accepted at Middleware’21

Weather and time influence the share of clean energy in the available power mix.
(Copyright: Unsplash)

Depending on weather conditions and electricity demand, the amount of emissions associated with the consumption of energy from the public power grid fluctuates over time, a metric which is commonly described as carbon intensity. The carbon intensity of the public grid can fluctuate highly during a single day: “In Germany, for example, it is not uncommon that during a sunny day at 1 pm a single kilowatt-hour emits less than 100 gCO2eq (grams of CO2-equivalent greenhouse gases), while the same kilowatt-hour consumed at 6 pm emits more than four times the amount, as the sun sets but demand rises”, explains Philipp Wiesner, PhD student at TU Berlin, working with the BIFOLD researchers Prof. Dr. Odej Kao and Prof. Dr. Lauritz Thamsen. Likewise, the emissions vary between different regions, averaging at 313 gCO2eq/kWh in Germany and only 56 gCO2eq/kWh in France, where most of the energy is generated from nuclear power plants.

Making cloud computing more sustainable

Shifting delay-tolerant computational workloads towards times where energy is expected to be clean, can reduce the associated carbon emissions significantly. In their publication Philipp Wiesner and his co-authors Ilja Behnke, Dominik Scheinert, Kordian Gontarska, and Lauritz Thamsen analyzed the potential for exploiting this variability by delaying non-urgent workloads towards times when energy is clean, with the goal to make cloud computing more sustainable. 

They analyzed the power grids of Germany, Great Britain, France and California in 2020 with regards to their carbon intensity. “We then experimentally evaluated different workload shifting scenarios to investigate the influence of time constraints, scheduling strategies, and the accuracy of carbon intensity forecasts”, says Philipp Wiesner. “In contrast to many other researchers, we did not focus on reducing the total amount of energy used but on using energy at the right time – which may vary from region to region.“

Scheduling workloads at times when the carbon intensity is expected to be low, can reduce the carbon footprint of data centers.
(Copyright: Philipp Wiesner)

Examples for energy-intensive but flexible workloads range from large machine learning jobs and scientific simulations to data processing pipelines and video renderings. A user that issues a machine learning job on a friday 6 pm may not care whether the job is done by 11 pm the same day or monday morning, as long as it is done by the time she returns to her desk to analyze the results.“ However, according to simulations conducted in the research paper, this additional flexibility can already reduce the job’s carbon emissions by 5.7 percent to 8.5 percent, as the carbon intensity is usually lower during weekends. Another category of flexible jobs are periodic batch jobs such as nightly compile jobs, integration tests, database backups or the generation of business reports. While current contracts between service providers and customers promise to always execute such jobs at certain times, they could also provide time windows. The increased flexibility could reduce emissions by up to a third in certain areas” summarizes Philipp Wiesner.

In the media:

New Scientist (November 01, 2021) : Smart scheduling for big computing tasks cuts emissions up to a third

The publication in detail:

Philipp Wiesner, Ilja Behnke, Dominik Scheinert, Kordian Gontarska, Lauritz Thamsen: Let’s Wait Awhile: How Temporal Workload Shifting Can Reduce Carbon Emissions in the Cloud. Middleware 2021, to appear
Preprint [PDF]

Abstract

Depending on energy sources and demand, the carbon intensity of the public power grid fluctuates over time. Exploiting this variability is an important factor in reducing the emissions caused by datacenters. However, regional differences in the availability of low-carbon energy sources make it hard to provide general best practices for when to consume electricity. Moreover, existing research in this domain focuses mostly on carbon-aware workload migration across geo-distributed data centers, or addresses demand response purely from the perspective of power grid stability and costs.
In this paper, we examine the potential impact of shifting computational workloads towards times where the energy supply is
expected to be less carbon-intensive. To this end, we identify characteristics of delay-tolerant workloads and analyze the potential
for temporal workload shifting in Germany, Great Britain, France, and California over the year 2020. Furthermore, we experimentally evaluate two workload shifting scenarios in a simulation to investigate the influence of time constraints, scheduling strategies, and the accuracy of carbon intensity forecasts. To accelerate research in the domain of carbon-aware computing and to support the evaluation of novel scheduling algorithms, our simulation framework and datasets are publicly available.

YouTube

By loading the video, you agree to YouTube’s privacy policy.
Learn more

Load video

Intelligent Machines Also Need Control

Home >

Intelligent Machines Also Need Control

Intelligent Machines Also Need Control

Dr. Marina Höhne, BIFOLD Junior Fellow, researches explainable artificial intelligence funded by the German Federal Ministry of Education and Research.

Happy to establish her own research group: Dr. Marina Höhne. (Copyright: Christian Kielmann)

For most people, the words mathematics, physics and programming in a single sentence would be reason enough to discreetly but swiftly change the subject. Not so for Dr. Marina Höhne, postdoctoral researcher at TU Berlin’s Machine Learning Group led by Professor Dr. Klaus-Robert Müller, as well as Junior Fellow at the Berlin Institute for the Foundations of Learning and Data (BIFOLD) and passionate mathematician. Since February 2020, the 34-year-old mother of a four-year-old son has been leading her own research group, Understandable Machine Intelligence (UMI Lab), funded by the Federal Ministry of Education and Research (BMBF).

In 2019, the BMBF published the call “Förderung von KI-Nachwuchswissenschaftlerinnen” which aims at increasing the number of qualified women in AI research in Germany and strengthening the influence of female researchers in this area long-term.
“The timing of the call was not ideal for me, as it came more or less right after one year of parental leave,” Höhne recalls. Nevertheless, she went ahead and submitted a detailed research proposal, which was approved. She was awarded two million euros funding over a period of four years, a sum comparable to a prestigious ERC Consolidator Grant. “For me, this came as an unexpected but wonderful opportunity to gain experience in organizing and leading research.”

A holistic understanding of AI models is needed

The topic of her research is explainable artificial intelligence (XAI). “My team focuses on different aspects of understanding AI models and their decisions. A good example of this is image recognition. Although it is now possible to identify the relevant areas in an image that contribute significantly to an AI system’s decision, i.e. whether the nose or the ear of a dog was influential in the model’s classification of the animal, there is still no single method that conclusively provides a holistic understanding of an AI model’s behavior. However, in order to be able to use AI models reliably in areas such as medicine or autonomous driving, where safety is important, we need transparent models. We need to know how the model behaves before we use it to minimize the risk of misbehavior,” says Marina Höhne outlining her research approach. Among other things, she and her research team developed explainable methods that use so-called Bayesian neural networks to obtain information about the uncertainties of decisions made by an AI system and then present this information in a way that is understandable for humans.

To achieve this, many different AI models are generated, each of which provides decisions based on slightly different parameterizations. All of these models are explained separately and subsequently pooled and displayed in a heat-map. Applied to image recognition, this means that the pixels of an image that contributed significantly to the decision of what it depicts, cat or dog, are strongly marked. The pixels that are only used by some models in reaching their decision, by contrast, are more faintly marked.

“Our findings could prove particularly useful in the area of diagnostics. For example, explanations with a high model certainty could help to identify tissue regions with the highest probability of cancer, speeding up diagnosis. Explanations with high model uncertainty, on the other hand, could be used for AI-based screening applications to reduce the risk of overlooking important information in a diagnostic process,” says Höhne.

Standup meeting: Each group member only has a few minutes to explain his or her scientific results.
(Copyright: Christian Kielmann)

Today, the team consists of three doctoral researchers and four student assistants. Marina Höhne, who in addition is associated professor at the University of Tromsø in Norway, explains that the hiring process of the team came with problems of a very particular nature: “My aim is to develop a diverse and heterogeneous team, partly to address the pronounced gender imbalance in machine learning. My job posting for the three PhD positions received twenty applications, all from men. At first, I was at a loss of what to do. Then I posted the jobs on Twitter to reach out to qualified women candidates. I’m still amazed at the response – around 70,000 people read this tweet and it was retweeted many times, so that in the end I had a diverse and qualified pool of applicants to choose from,” Höhne recalls. She finally appointed two women and one man. Höhne knows all about how difficult it can still be for women to combine career and family. At the time of her doctoral defense, she was nine-months pregnant and recalls: “I had been wrestling for some time with the decision to either take a break or complete my doctorate. In the end, I decided on the latter.” Her decision proved a good one as she completed her doctorate with “summa cum laude” while also increasing her awareness of the issue of gender parity in academia.

Understandable AI combined with exciting applications

Höhne already knew which path she wanted to pursue at the start of her master’s program in Technomathematics. “I was immediately won over by Klaus-Robert Müller’s lecture on machine learning,” she recalls. She began working in the group as a student assistant during her master’s program, making a seamless transition to her doctorate. “I did my doctorate through an industry cooperation with the Otto Bock company, working first in Vienna for two years and then at TU Berlin. One of the areas I focused on was developing an algorithm to make it easier for prosthesis users to adjust quickly and effectively to motion sequences after each new fitting,” says Höhne.  After the enriching experience of working directly with patients, she returned to more foundational research on machine learning at TU Berlin. “Understandable artificial intelligence, combined with exciting applications such as medical diagnostics and climate research – that is my passion. When I am seated in front of my programs and formulas, then it’s like I am in a tunnel – I don’t see or hear anything else.”

Marina Höhne has a passion for math. (Copyright: Christian Kielmann)
More information:

Dr. Marina Höhne
Understandable Machine Intelligence Lab (UMI)
E-Mail: marina.hoehne@tu-berlin.de

Award for paper on processing semantic data streams

Home >

Award for paper on processing semantic data streams

Award for paper on processing semantic data streams

Congratulations to BIFOLD Fellow Prof. Dr. Manfred Hauswirth and BIFOLD Junior Fellow Dr. Danh Le-Phuoc: At the International Semantic Web Conference 2021 (ISWC-2021), their paper “A Native and Adaptive Approach for Unified Processing of Linked Streams and Linked Data” achieved 2nd place for the “SWSA Ten-Year Award” with an honourable mention.

The paper by Danh Le-Phuoc, Minh Dao-Tran, Josiane Xavier Parreira, and Manfred Hauswirth presents one of the first systems for efficient, highly scalable processing of structured, semantic data streams (dynamic data) in combination with background knowledge (static data). The performance of the system was the best among the systems available at the time of the publication in 2011, it has significantly shaped and influenced the research field of semantic data stream processing, and continues to have an impact to the present day. The system makes it possible to easily and efficiently build applications in many fields, e.g. traffic monitoring, environment, Internet-of-things, edge computing, and the smart city.

Dr. Danh Le-Phuoc
Copyright: BIFOLD

“We are very proud that our paper is recognized by one of the leading conferences on semantic web as one of the most impactful papers. At BIFOLD we continue to work on the CQELS Framework, a semantic stream processing and reasoning framework for developing semantic-driven stream and event processing engines on both edge devices and cloud.”

The International Semantic Web Conference is one of the most important conferences in this research area. Every year at the event, the most influential contributions from the conference proceedings of the last ten years are honored by the Semantic Web Science Association with the “SWSA Ten-Year Award”. The decision is based primarily, but not exclusively, on the number of citations received by the papers from the conference proceedings during the past decade. 

The publication in detail:

Danh Le Phuoc, Minh Dao-Tran, Josiane Xavier Parreira, Manfred Hauswirth: A Native and Adaptive Approach for Unified Processing of Linked Streams and Linked Data. ISWC (1) 2011: 370-388
[PDF]

Abstract

In this paper we address the problem of scalable, native and adaptive query processing over Linked Stream Data integrated with Linked Data. Linked Stream Data consists of data generated by stream sources, e.g., sensors, enriched with semantic descriptions, following the standards proposed for Linked Data. This enables the integration of stream data with Linked Data collections and facilitates a wide range of novel applications. Currently available systems use a “black box” approach which delegates the processing to other engines such as stream/event processing engines and SPARQL query processors by translating to their provided languages. As the experimental results described in this paper show, the need for query translation and data transformation, as well as the lack of full control over the query execution, pose major drawbacks in terms of efficiency. To remedy these drawbacks, we present CQELS (Continuous Query Evaluation over Linked Streams), a native and adaptive query processor for unified query processing over Linked Stream Data and Linked Data. In contrast to the existing systems, CQELS uses a “white box” approach and implements the required query operators natively to avoid the overhead and limitations of closed system regimes. CQELS provides a flexible query execution framework with the query processor dynamically adapting to the changes in the input data. During query execution, it continuously reorders operators according to some heuristics to achieve improved query execution in terms of delay and complexity. Moreover, external disk access on large Linked Data collections is reduced with the use of data encoding and caching of intermediate query results. To demonstrate the efficiency of our approach, we present extensive experimental performance evaluations in terms of query execution time, under varied query types, dataset sizes, and number of parallel queries. These results show that CQELS outperforms related approaches by orders of magnitude.

New Berlin Cell Hospital announced

Home >

New Berlin Cell Hospital announced

New Berlin Cell Hospital announced

When cells make the wrong decision, diseases ensue. This insight came from Berlin – namely from Rudolf Virchow. On October 13, 2021, at an event celebrating the 200th birthday of the famous pathologist, physician and socialist politician Rudolf Virchow, the Max Delbrück Center for Molecular Medicine in the Helmholtz Association (MDC) and the Charité – Universitätsmedizin Berlin, together with several Berlin research institutions, declared the founding of the Berlin Cell Hospital, an Institution to shape and develop the cell-based medicine of the future.

From left to right: Prof. David Horst (Charité), Prof. Heike Graßmann (MDC), Dr. Stan Gorski (MDC),
Prof. Nikolaus Rajewsky (MDC), Prof. Christopher Baum (Charité), Prof. Otmar D. Wiestler (Helmholtz-Gemeinschaft),
Prof. Angelika Eggert (Charité), Prof. Heyo K. Kroemer (Charité), Michael Müller (Regierender Bürgermeister Berlin)
and Prof. Thomas Sommer (MDC)
(Copyright: Charité / Sabine Gudath)

The Berlin Cell Hospital will begin life as a registered association to allow established institutions to become members. It’s key participants are the MDC, the Helmholtz Association, Charité, the Berlin Institute of Health at Charité (BIH) and the Berlin Institute for the Foundations of Learning and Data (BIFOLD). The Cell Hospital also hopes to cooperate with private partners and other institutions in Berlin and Germany, including the Helmholtz Health Centers and the German Centers for Health Research (DZGs), as well as to create an international network.

The number of chronically ill people who require expensive and invasive treatments is growing continuously. At the same time, life expectancy is on the rise, which means the population is getting older and older. Instead of only treating common diseases when their patients start to display serious symptoms – by which time a great deal of damage has already been done – doctors are in urgent need of new diagnostic and therapeutic strategies.

Diseases often begin much earlier than the onset of symptoms. As far back as 1858, the famous pathologist Rudolf Virchow suggested that the origin of diseases can be traced back to changes in individual cells. So how and why do these changes occur?

Each cell is continuously “reading” the genome so it knows how to react to signals from neighboring cells or new environmental conditions. How exactly each individual cell interprets this “book of life,” but also what mistakes happen in the process and which changes disrupt the process, is something scientists have only been able to observe for the past few years thanks to single-cell biology. The volume of data generated for each cell corresponds in magnitude to that produced by classical genomics techniques. The amount of information it contains is unimaginable, and the depth of detail unparalleled. BIFOLD will contribute machine learning research and tools that make this flood of  data manageable.

The Berlin Cell Hospital brings together experts from clinical practice, biomedical research, technology, data science, mathematics and engineering science.
(Copyright: Unsplash)

“It’s as if we discovered a super microscope,” says Nikolaus Rajewsky. “Thanks to these technologies, we can analyze every single cell in a tissue for the very first time and understand when and why it gets sick.” Cell-based medicine wants to use this knowledge to guide cells back to a state of health as quickly as possible – with the help of extremely early diagnostics that recognize when a cell takes its first step in the direction of disease, with the help of targeted procedures on molecular mechanisms and with cellular therapies, RNA-based approaches and similar techniques. The goal of cell-based medicine is to close the gap between classic prevention and medicine that treats only symptomatic patients. Thanks to its personalized treatment strategies, the concept is also suitable for preventing disease relapses and resistance to immunotherapy or chemotherapy.

But successfully implementing cell-based medicine is no easy feat. It requires a multifaceted approach that breaks down disciplinary and institutional boundaries and that, to date, has never existed under one roof in Germany. To understand diseases in a new way, a research concept that brings together experts from clinical practice, biomedical research, technology, data science, mathematics and engineering science is needed – all working together in close proximity to advance novel approaches to medicine. The core pillars are single-cell technologies, patient-specific model systems such as organoids, and new AI solutions. In the new Cell Hospital, these will mainly be applied to the major chronic diseases (cancer, cardiovascular diseases, infectious diseases and neurological diseases).

The Cell Hospital aims to develop molecular prevention strategies and new precision diagnostics, as well as to reliably identify new drug targets for molecular and cellular therapies. In order to transfer knowledge as quickly as possible, the Berlin Cell Hospital is planning a broad innovation and industry program – for example via the Virchow 2.0 Clusters4Future application – which should facilitate dynamic developments and remove any existing obstacles. The resulting innovation ecosystem will hopefully include industry partnerships, cross-sector networking, innovation spaces and labs, and should promote a thriving spin-off culture. An education and training program will target the workforce of the future, including students, researchers and health professionals. The Berlin Cell Hospital aims to unite these basic elements and the critical mass of science under one roof – in close proximity to the clinic and patients, and in a way that involves patients and citizens from the very beginning.

In Search for Algorithmic Fairness

Home >

In Search for Algorithmic Fairness

In Search for Algorithmic Fairness

Artificial intelligence (AI) has found its way into many work routines – be it the development of hiring procedures, the granting of loans, or even law enforcement. However, the machine learning (ML) systems behind these procedures repeatedly attract attention by distorting results or even discriminating against people on the basis of gender or race. “Accuracy is one essential factor of machine learning models, but fairness and robustness are at least as important,” knows Felix Neutatz, a BIFOLD doctoral student in the group of Prof. Dr. Ziawasch Abedjan, BIFOLD researcher and former professor at TU Berlin who recently moved to Leibniz Universität Hannover. Together with Ricardo Salazar Diaz they published “Automated Feature Engineering for Algorithmic Fairness“, a paper on fairness of machine learning models in Proceedings of the VLDB Endowment.

Statue of Lady justice
BIFOLD researchers suggest a new machine learning model that leads to both: high accuracy and fairness.
(Copyright: Pixabay)

Algorithms might reinforce biases against groups of people that have been historically discriminated against. Examples include gender bias in machine learning applications on online advertising or recruitment procedures.

The paper presented at VLDB 2021 specifically considers algorithmic fairness. “Previous machine learning models for hiring procedures, usually discriminate systematically against women”, knows Felix Neutatz: “Why? Because they learn on old datasets derived from times when fewer women were employed.” Currently, there are several ways to improve the fairness of such algorithmic decisions. One is to specify that attributes such as gender, race or age are not to be considered in the decision. However, it turns out that other attributes also allow conclusions to be drawn about these sensitive characteristics.

The state-of-the-art bias reduction algorithms simply drop sensitive features and create new artificial non-sensitive instances to counterbalance the loss in the dataset. In case of recruiting procedures, this would mean simply adding lots of artificially generated data from hypothetical female employees to the training dataset. While this approach successfully removes bias it might lead to fairness overfitting and is likely to influence the classification accuracy because of potential information loss.

Felix Neutatz.
(Copyright: Privat)

“There are several important metrics that determine the quality of machine learning models,” Felix Neutatz knows, “these include, for example, privacy, robustness to external attacks, interpretability, and also fairness. The goal of our research is to automatically influence and balance these metrics.”

The researchers developed a new approach that addresses the problem with a feature-wise, strategy. “To achieve both, high accuracy and fairness, we propose to extract as much unbiased information as possible from all features using feature construction (FC) methods that apply non-linear transformations. We use FC first to generate more possible candidate features and then drop sensitive features and optimize for fairness and accuracy”, explains Felix Neutatz. “If we stick to the example of the hiring process, each employee has different attributes depending on the dataset, such as gender, age, experience, education level, hobbies, etc. We generate many new attributes from these real attributes by a large number of transformations.  For example, such a new attribute is generated by dividing age by gender or multiplying experience by education level. We show that we can extract unbiased information from biased features by applying human-understandable transformations.”

Finding a unique feature set that optimizes the trade-off between fairness and accuracy is challenging. In their paper, the researchers not only demonstrated a way to extract unbiased information from biased features. They also propose an approach where the ML system and the user collaborate to balance the trade-off between accuracy and fairness and validate this approach by a series of experiments on known datasets.

The publication in detail:

Ricardo Salazar, Felix Neutatz, Ziawasch Abedjan: Automated Feature
Engineering for Algorithmic Fairness
.
PVLDB 14(9): 1694 – 1702 (2021).

Abstract

One of the fundamental problems of machine ethics is to avoid the
perpetuation and amplification of discrimination through machine
learning applications. In particular, it is desired to exclude the influence of attributes with sensitive information, such as gender or
race, and other causally related attributes on the machine learning
task. The state-of-the-art bias reduction algorithm Capuchin breaks
the causality chain of such attributes by adding and removing tuples. However, this horizontal approach can be considered invasive
because it changes the data distribution. A vertical approach would
be to prune sensitive features entirely. While this would ensure fairness without tampering with the data, it could also hurt the machine learning accuracy. Therefore, we propose a novel multi-objective feature selection strategy that leverages feature construction to generate more features that lead to both high accuracy and fairness.
On three well-known datasets, our system achieves higher accuracy
than other fairness-aware approaches while maintaining similar or
higher fairness.