Rise of AI Summit 2020 · Berlin · Applied AI Day | Future Space

Trustworthy AI for the Physical World

Ulli Waltinger · Technology Head, Siemens AI Lab

At Rise of AI Summit 2020 in Berlin, Ulli Waltinger, Technology Head of the Siemens AI Lab, argues that industrial AI has moved past the question of whether machine learning adds value and must now be put into pervasive operation across energy management, mobility, additive manufacturing and shop-floor systems. He distinguishes two application classes: 'selfish systems' that increase machine autonomy, and 'digital companions' that support human decision-making. Because labeling all industrial data is infeasible, Waltinger says the field is shifting from purely supervised learning toward self-supervision combined with domain expertise. He stresses starting from the deployed business problem and the 'minimal justifiable improvement' rather than chasing proof-of-concept accuracy, and names data readiness, active learning, explainability, federated learning and robustness against domain shift as the core applied-research questions his lab works on with partners. He closes with three takeaways: aim for production, not prototypes; define the minimal justifiable improvement and what not to do; and treat trustworthiness, explainability, robustness and security as imperative for industrial AI.

Key takeaways

  • Industrial AI splits into 'selfish systems' that raise machine autonomy and 'digital companions' that support human decisions.
  • Because most industrial data cannot be labeled at scale, Waltinger's lab favors self-supervision combined with human domain expertise.
  • Projects should start from the 'minimal justifiable improvement' needed in a real business process, not from available data or accuracy.
  • Waltinger proposes an 'AI Ops twin' that tests models in simulation before deploying them to production devices.
  • There is no unbiased dataset; bias can only be monitored continuously, not eliminated with a single tool.

Detailed notes

From selfish systems to digital companions

Waltinger frames industrial AI around two application classes: 'selfish systems', which increase machine autonomy for adaptive, cost-saving operation, and 'digital companions', which embed machine intelligence into human decision-support systems. Siemens, he says, connects a large installed base of physical assets such as turbines, trains, compressors and wind turbines, bridging the virtual and physical worlds. Because not all data can be labeled at scale, he argues the field is moving from a purely supervised, data-driven paradigm toward a 'system-driven' one that combines self-supervision with human domain expertise, especially in industrial settings where data is often unlabeled or ambiguous (is a sensor reading an outlier or an anomaly).

Start from the deployed problem, not the data

Waltinger says research in industrial settings works backward from research (fix the data, tune the model) to application (fix the model, tune the data), driven by business impact rather than technical novelty. He introduces the 'minimal justifiable improvement' as the guiding question: what is the smallest performance gain that actually creates value in a given dashboard or decision process, and in which business process must it be integrated. He argues teams should start from the deployed use case and work backward, rather than beginning with a proof of concept, which he says tends to end in a 'PoC cycle' that never reaches production. Data readiness (an organization's maturity in accessing, sharing and using its own data) is a second recurring obstacle he sees across corporate partners.

Research directions: active learning, knowledge graphs, robustness

Waltinger lists four recurring research questions from his lab's work with partners over the past three years: data-efficient learning (active learning, simulation, prior-knowledge induction) where not every data asset is connected; multi-modal interaction for companion systems deployed quickly and improved continuously; explainability and federated learning across distributed physical assets; and, as the focus for the next two years, trustworthy predictions and reliable operation, including safety architectures and robustness. He notes that softmax output layers can behave erratically outside the training distribution, and that industrial deployments demand transfer across different cameras, angles and setups, pushing toward robustness and domain generalization. Active learning and annotation, he says, let domain experts inject their own knowledge and flag what counts as an anomaly. Knowledge graphs are described as a useful, explainable way to digitize domain structure and bridge internal and external data.

Shipping and the AI Ops twin

In operation, Waltinger favors architectures that are 'tightly integrated but loosely coupled', combining open source with proprietary on-premises software rather than re-patching open source components. He proposes an 'AI Ops twin': testing and deploying machine learning models in a simulated environment before targeting production devices, analogous to process and product simulation. He also stresses team structure, arguing against relying on a single generalist ('the one-trick pony') and for bringing together data science, operations and business expertise, and against designing one-off solutions instead of repeatable processes. He credits a colleague, Diego, with the point that a machine learning project without an executive sponsor rarely succeeds.

Bias, explainability and key takeaways

Waltinger rejects the idea of an 'unbiased button': data is always biased, and tooling can only reveal distribution and coverage gaps, not remove bias outright; monitoring must be continuous, which is part of why he calls the deployment loop an AI Ops twin. On explainability, he says neural networks are not inherently full black boxes; transparency about how a decision was made, how a model was trained and how it behaves post-deployment can be engineered, sometimes by combining explainable and non-explainable models. His three takeaways: favor production over easy prototyping; define the minimal justifiable improvement and explicitly what not to do, to focus limited resources; and treat trustworthiness, explainability, robustness and security as imperative for industrial-grade AI.

Q&A highlights

Asked whether demanding business-driven AI decisions risks killing experimentation, Waltinger says the lesson from the 'big data hype' was that management started demanding double-digit business cases up front and that this killed projects; he argues AI will touch every process and product eventually, so the real question is whether to start now or later, and recommends starting from the minimum justifiable improvement rather than treating proof-of-concept testing, which he compares to entering an ImageNet competition, as an end in itself.

Quotes

Machine learning can be truly a fragile beast.

Ulli Waltinger

There is no unbiased buttons, from my perspective.

Ulli Waltinger

Design for the deployment, design the minimal justifiable improvement, design for that.

Ulli Waltinger

Questions answered in this talk

What does Ulli Waltinger mean by 'minimal justifiable improvement' in industrial AI?

According to Ulli Waltinger at Rise of AI Summit 2020, it is the smallest, clearly defined performance gain a machine learning application needs to deliver real value in a specific business process, and it should be identified before deciding how to build or deploy the model.

Why does Ulli Waltinger say there is no such thing as unbiased data?

According to Ulli Waltinger at Rise of AI Summit 2020, all real-world data carries some bias, and tools can only reveal coverage and distribution gaps and support continuous monitoring, not remove bias with a single 'unbiased button'.

What is an 'AI Ops twin' as described by Ulli Waltinger?

According to Ulli Waltinger at Rise of AI Summit 2020, it is a practice of testing and deploying machine learning models in a simulated virtual environment before applying them to the target physical device, mirroring process and product simulation.

Mentioned

Minimal justifiable improvement, Active learning, Knowledge graphs, Federated learning, Domain generalization, AI Ops twin, Explainability, Data readiness

Keep exploring

More from Ulli Waltinger

More from Rise of AI Summit 2020

More on Enterprise AI & Adoption

Transcript

Auto-generated transcript, not edited. · 5,037 words

Show the full transcript

i'll introduce our speaker again of course he's deeply interested in the foundations of computational intelligence and he's a zealot when it comes to an open and interdisciplinary research and innovation culture he has more than 40 publications and more than 100 talks under his belt he's the technology head of the siemens artificial intelligence lab in munich that he founded with his team in 2017 and i hope that he hears me and is here uli waltenga thanks very much yes i can hear you can you hear me okay yes i can hear you too i see a slide and i hope that's what i what we're supposed to see right now exactly so uh first of all congratulations you know this team behind the scene and from the scene you did an amazing job i super enjoyed the rise of ai event amazing event thanks very much and i'm more than honored folks out there um to talk to you a bit on trustworthy ai in the physical world actually uh my name is julie barting and

i pretend work here from from a lot of teams actually out there i'm actually working for one of the older startups i guess in berlin isn't it right uh so siemens and um if we're talking about ai and we're talking about industrial eye i guess we are now at the point where you know we need to put industrial ai in operation right tower a focus and it's ours a pervasive use and if you look in a large corporation such as i'm devoted to work for right um um you know as broad that the applications are from energy management you know building grid controllers from putting more autonomy in mobility streams from what it means actually you know bridging the world between simulation and machine learning what it means actually control optimization obviously or you know bringing and shipping additive manufacturing to the next level you know or robotic systems that you know um it can be more adaptive more independent um to new new demands but obviously also in in the shop floor area right

more throughput more availability you know all these kinds of tracks and we along the line of of these use cases uh you know which need to provide some value right um we can somehow you know discriminate between two different application areas the one is the natural reflex i guess selfish systems what we call that means like we need to increase the level of machine autonomy that means you know we want to have adapts and a breath of value and more adaptive more efficient right cost saving these kinds of adaptive um uh technologies um you know uh on on the horizons and the second is i guess the natural efflux as well right it's assistance right we called it digital companions so embodying some kinds of capabilities along the line of machine intelligence in certain decision support systems so autonomy and assistance are maybe the leading principles that we see in the industrial area second obviously you know our job industry i mean you know we connect the physical individual world and this is i i

guess um the interesting parts in there right uh we work with partners which have a a whole you know install base of industrial assets coming from um turbines coming from trains coming from compressors come from wind turbines obviously as well and um the challenge is you know how do we connect this asset together and how do we make sure that the virtual world and the physical world are bridging together and obviously data does play a role right we are nonetheless obviously in a data-driven society somehow in this let's say supervised data-driven society isn't it right um but it's also clear that you know um we will not have the time to actually label all the data in our world that is accessible or not yet accessible right so we will move towards a what we call the system driven you know um society more so relying more on self-supervision relying more on that not every data is being sensed automatically at the same time we rely heavily especially in industrial flavors on expertise on domain

or how and so human driven systems and i guess you know those worlds bring together the implicit and hybrid knowledge but also the aspects moving further away from the purely supervised let's say weak machine learning paradigms and what helps here and i think i love yesterday actually they call like uh ecosystem beats um you know um ecosystem beats a strategy love that um and i think you know the ecosystem of applied research is the core and the key actually how to accelerate further right because applied research is for me the impact driven research you're not not only advancing state of the art obviously right but actually applying and and you know challenging real world challenges and and helping enabling you know to transform these ideas into innovations and from a geeky perspective we can say you know in research we we you know we fix the data and we tune the model somehow right we advance knowledge and we find good architectures once we you know identified a great architecture great model um it comes

to application and in application it might be a vice versa right we fix the model and tune the data so we apply different data and different applications to a given you know computer vision architecture predictive maintenance architecture rv detection and so on and the business impact is driving and defining um actually these these questions this research question employee and it ain't that easy you know what's a different impact what is what is does provide value we are in current stage far already far far further that you know is machine learning providing any value that's it's not it's a no-brainer right it's you know it's the questions what is actually i love uh you know the notion of the minimal justifiable improvement what is the right problem to tackle and this is how you're actually starting right from where why to how to what what is the right problem what are the right constraints what number needs to pop up in which dashboard in which axle sheet in in which the schism decision making uh you

know application um that gets you your bony this is the problem this is this is something wagens where does it actually be integrated so identifying you know what is good enough you know is it one percent is it two percent right what are the key drivers of the performance right um this is such an essential part in identifying it in first place don't worry about data in the first place worry about what is the right problem then the second step is how in which business process it needs to be integrated right and once you have that you know then you can talk okay what what do we need to do in order to enable that right so start from the deployed object from the deployed object and work backwards not the other way around otherwise you are ending up in the proof right in the proof of concept cycle you don't want to have second obviously you know data data readiness if you if i talk to partners you know all around the corporates right

they all suffer data and i mean with data rather than us hey for sure right it's not a yeah we have some data in some cop europe or asset right data readiness means you are mature enough in access in delivering in sharing and utilizing your data right and there's a whole spectrum actually you know what can be considered there you know talk to us or talk to me if you you know think about hey how do we address data readiness in your cooperation and third obviously is about research questions isn't it right and really real world challenges you know drive the applied research direction actually have and these are you know sometimes the blueprints that we use it throughout i don't know the last three years and i guess we'll use it also for the next three years right what's the some typical questions we have and we ask for the you know applied research ecosystems and the partners out there is you know how do we do with data efficient learning right not every

data asset you know is being connected is available right don't you know what's the role of active learning what's the role of simulation how do we do prior knowledge induction second about companion you know what is multi-modal interaction what is if we depl deploy very fastly end-to-end and then continuously increase on the accuracy and the level with together with apartments third pillar obviously you know explainability is something which especially in physical domains and physical words is quite important but also about you know the capabilities of for instance federated learning right can we distribute can we learn from a distributed asset base nice and fourth obviously and i think this will be the next two years of you know trustworthy predictions and reliable reliable operations so we need to think about safety architectures about robustness but also about what it means industrial trustify great solutions so trust seems to be a bit of a role right right so in in simple words right you we need to talk about industrial grade transverse solutions to be honest

right because um obviously you have you know the aki capabilities and i love the accessibility of current ai technology right starting a tensorflow or pytorch getting the pi you know notebook up and running you know easy nice done in maybe a couple of hours right for most of itself but what it means actually providing value and to and is a bit of a different twist right because in order to solve the right things you it requires some kind of understanding you know in what kind of business integration needs to be the second is that if it changes the business process right it needs to be deployed in a productive system and this requires trust if the system automatically triggers some kind of changes in the business process and obviously since we don't don't want to have this 101 and we have a continuous loop actually in deploying it it requires automation on the machine learning capabilities and this brings us together the industrial grade trustworthy solutions a bit more than launching a python new jupyter

notebook especially in this industrial ai you know the challenge is that we need to make from both worlds actually some some benefits the implicit data obviously data driven you know inside but also the explicit stuff domain know how it integrated because the data you seen in the industrial flare you know it's not easy to label them right it may be cats and dogs for the other world but sensor is it outlier is it animally right a lot of data is not yet sensed so brownfield environments you know the coke and there's a lot of domain or how because you know we we need to have some sense of feeling how is actually moving how is the physical control how is the actor actually been moving so that brings a bit of some challenges in out there right and we all know from our daily i guess you know consumer tech experience machine learning can be truly a fragile beast right uh obviously you know in certain aspects and for most of the aspects if you

interface on a daily basis with machine learning technology we don't necessarily expect a hundred percent performance it's okay if it's ranked two or if you bring you know well potentially not a wrap alpaca you know to your wife because somebody recommended in the store that you know there's a secret replacement for that right it's a bit of fragile right and this i guess you know respect the past but you know invent the future correlation is not causation right causality is for for sure reasons you know the last two years one of the the key pillars of research right in the machine learning departments right and i guess it summarizes also like dad nigel i love that because this is something how it feels in in many cases right french is a crazy and a beautiful language you know it's it's spelled different than you speak it right the sound of the language is magnificent but it's somehow a bit of a difference but still you know machine learning it looks like that right you could

just you know mesmerizing everything memorizing everything and that's still you know cracking up on the french travel championship without speaking actually any french isn't it and for for business what does it actually mean for us we demand scalability solutions that means for sure you can start the poc in certain factory right you can train a model in a certain and apply it in a certain applications but we actually demand scalable solutions that means we need to rely on that if we train certain solutions we can transfer them across different deployments and that means we need to cope with environmental discrepancies but also deployment discrepancy different cameras different angles different data different setup you know and this requires quite some push towards robustness and quite some push towers what is called domain generalization and this is where we also start together with university with the partners but also in our own turf right it says you know how do we make sure that you know we we advance the field of trustworthy predictions right um if

you use half max you know network so you have the most of the layers right at the end you use the soft mix layers for get some distribution if you put that out of the main it does crazy things guys right isn't it so how do we make sure that we you know get some feedback right um on the robustness of the applications if you move the machine learning model from one to the other at the same time obviously as we as we build up this this large large nets right on there certain capabilities right we need to deploy it in the physical world somewhere on the edge level and the edge may be a sparse resource edge right a sparse resource device which you not consume like you know megabytes um to inference nevertheless right we need to provide reliable models with a low memory footprint you know for years to run um it's also a different question right at the same time you know obviously you know how do we approach that you

know we're not seeing this hundred thousands of events uh that we have in the consumer space on the ad rate but how do we actually use a matter learning uh mechanisms in order to come with a few gradient steps few shot analysis you know to sufficient detection algorithms right this is all typical problems that arise in the industrial domain which we approach and which we also share with the research and academia partners out there next level is obviously active learning active learning helps us primarily also obviously getting feedback done that means from annotation obviously going fast and end-to-end from data to actually you know actuation but allowing the experts and the partners to to seeing and listening but also injecting their own knowledge this is an anomaly this is you know to forget about it this is interest this is not seen yet right and active learning and annotation gives us here you know on two sides the domain of capturing process as well as you know enforcing a bit more explainability and i guess

we have a lot you know heard yesterday already right from from hans and so on you know but knowledge graphs right yeah for sure we are in a huge height of data management again right remember the big data hype um and yes knowledge graphs are interesting and very very you know great and easy to go actually way in digitizing these kinds of domain structures you know digitizing uh information breaching silos but it allows us also you know to somehow you know enter the phase of graph based machine learning which is also very nice because it has also uh the explainability aspects in there so even more explicitly and it allows to inject external data to our internal data cells so that's quite quite interesting aspects but nevertheless you know once you have the methodologies and you work with your partners in there you know solving certain kinds of of pieces and then it's it's shipping it's you know in operation right and in operation is one of the you know paradigms is you know tightly

integrated but loosely coupled because open source yes plays a vital role right in industrial eye also on the shop floor level or in corporate level but there will be also always proprietary software you need to eat again and we don't want to recharge again with this patching of open source you know patches um that you know as are adopted somehow as we you know the best practice learned from uh from obviously the you know big data hype right so it's it's the main idea is highly integral but loosely coupled there will be always on-prem stuff going on right and that brings actually a new horizon which i would say is is super hard and currently um being you know put together it's it's the notion of you know that there is an ai ops the industrial ai ops twin actually as it's a few of deployment and tested in virtual environments but then also to target to the device of production we applied similar aspects in the simulation domain process simulation product simulation you know

having a virtual representation before it's actually been constructed in in some kind of plants and i think we need exactly these kinds of twin in an ai ops direction as well we need to apply certain machine learning capabilities in a simulated in a virtual environment before we go actually on the target devices um given this here which we call the that upstream and there is many reason right why somehow enter and and the pocs and we had uh this morning quite quite feedback already right uh why is in i think it's you know one of the aspects is you know most of the companies i see you know they design proof of concepts don't don't don't go for proof of concept right i think we are over that stage right design for the deployment design the minimal justifiable improvement design for that right second obviously team structures right every company has this you know our working students does that he's the ai ops he's the business analyst he's the software engineer he's you know this

one trick pony right i think don't do it right bring people together ops needs different expertise from data science to operationalize from business to customer right it needs different capabilities to make use of your team that you mostly have or use the partners out there second a third one is obviously you know technology technology you know we need best practices obviously but we need also to design for repeat repeat certain aspects it should be not the one thing you know we did it once but it's always you know focus on is that repeatable in the process and last but not least i think kudos to diego here right he said like you know machine learning project without an executive sponsors you know you know never see the light you know and i think that's that's some truly thing a lot of people bring the opportunity go for the plc but if you don't have a really a corporate sponsoring order you know this decision maker in the you know aligned with his values interests right

what is we need to do in order that you get your boning right this this brings value statement sometimes on their ends you know this helps obviously as well you know shifting closer to and to end integration and last but not least right i think we are we are talking a lot of data and um i've seen the first pitches already coming on with you know i we have this magnificent tool you know of this unbiased button you know what the heck right there is no unbiased buttons from my perspective right we deal with buyers and and data we deal with data and data is dirty and you know bias as it is right we can have a closer look we cannot monitor but you know it will be always somehow flawed in this world somehow by us by certain aspects and this brings us actually to an aspect that's you know it's it's it's um that we we need to reflect on on the actually applications that we built that are fueled by certain

technologies fueled by certain uh data right um which obviously comes with a lot of capabilities that we have right like visual monitoring allows some transparency it allows some security something pretty good right or you know process efficiency or you know um carson failure reductions in certain capabilities but sometimes these you know use cases bring also some accompanied by some kind of risks right like you know misuse or bias or you know safety you know or you know in transparencies or black boxes right but there are certain mitigation scenarios one and as you as a decision maker can take right whether it's awareness whether it's training whether it's technology it's where it's co-creation maybe sometimes also regulations right in helping to mitigate them but i think c-level needs some kind of blueprints and a playbook and in order not only seeing capabilities but also not only seeing risks but actually you know having a playbook of awareness how to approach certain use cases um enabling and mitigating risks and this brings me actually to the to

the key takeaways i guess right and the famous three right aim models obviously you know easy to prototype it's super super nice you know the accessibility of machine learning methodology i love it right um but it's a different if you maintain it in production but you go for enter and this is you know even though you're not having the highest accuracy continuous improvement needs to be targeted but it needs to go enter and otherwise you don't do that second i'm a geek i love tac i love exploring technology but favor for business trip innovation what is the minimum justifiable you know improvements you want to have and define what you don't do right especially in large corporates right um you know there is a tendency of not defining what you don't do because if you know exactly what you don't do you can leverage this awesome you know ecosystem of which has the speed and scale that you actually need and third pillar you know is trustworthiness explainability robustness but also security you know will

be the imperative in industrial ai right and i'd love to engage with you further beyond ping me ping me on on this on the books or on the email or on linkedin more than happy to engage with you together thanks very much oli thank you so much uh i i think i have a couple of questions for you already although we only have a couple of minutes time for them but i was quite curious you dare to say there is no unbiased data right and i think that made many people are quite nervous i think because they're striving for unbiased data maybe some people in their companies told us no no of course we can remove bias from data on the other hand um you were talking about explainability and i think there are some people who would say there is also not a completely explainable ai what do you say to that yeah um let's take a look buy is first so you know bias is is there's always bias if you know if

you i think there's a great tooling in there um which helps sheds a light how is your distribution of the data yeah how is the coverage is it only is it only dogs or is there are there some random cats in there as well right stuff like that or you know how do we is what kind of features are being used right is gender included right or is gender explicitly excluded right um and it's but it also goes in the direction once even so if you have in the training material you know and if you're training set of of smart applications being said like okay i think it's a bit of a representative what we actually target then it hits the real world and different data comes in there right and so um i think you know if you talk about you're reflecting on the bias or on on the fragility of machine learning modules it's also clear that it doesn't stop with just the training but it's a continuous and that's why we call

it you know the ai ops twin it's a continuous monitoring what's going on here and to make sure you know we're aware of the implications that it can have second aspect explainability right and yeah for sure there was quite some some hip and there like neural networks all these black boxes you know and i don't know get what the what it's in there and stuff like that's why we don't use them yeah there are tools and technologies around there and there are certain um obviously pieces you know where a transparency and this is i guess what most demands in terms of explainability i totally agree with explainability but transparency right um i have currently no clue what my wife is doing i surely trust it i trust her we will do okay i don't have transparency but what she's currently doing but um in certain pieces right we need to invest and say like why certain decisions are being made and there are different versions and different actually instruments we can enforce a bit more

explainable features what it has learned how it behaves postmortem how it was trained how it's applied um but also maybe you know combining different explainable and non-explainable model in heaven a bit more closer so i don't see it as you know it's if you're using neural networks everything is black box it's ain't that so i i i there are measures to you know work together right thank you you were making a strong case for uh making business driven ai decisions and not tech driven ai decisions but at the same time we have been talking a lot about the proof of concept graveyard right i i think i i presume that probably the decisions to do a proof of concept are all tech driven right people want to experience want to see want to see if anything goes and because there is no business driving business decisions driving it then the the whole thing stops but is it really a good thing to demand business driven decisions because maybe people will just stop developing ai solutions

then because they don't see the business case yet they are not maybe they are not confident enough that they will have a successful end-to-end integration what do you think yeah but i i still go for that so um obviously you know lessons learned after the big data hype i don't know you remember that we had we had some cases there right some investments going there so the the the the natural reflex of of management was uh calculate me the business case if it's not not making a double-digit impact right we aren't going to go there right and that killed everything right and now they actually um pull down and say like yeah he's here's toy go it right but um you know um ain't gonna have a business impact we need to understand to the mindset right this is this is not not a totally magical tool but um if if you're thinking on the ai first it has a pervasive use in various that the quantity of applications right will bring the entire business

to together and and that's why you know i still urge sea level you need to understand this is not a holy grail but it will any different process in any given product there will be not a in there right so you want to start now or you want to start later this is the only question right and if we start now let's talk about the minimum justifiable improvement you want to make what does a certain application need to move that you get your bony right this is the question to ask and then you work backwards and i think that helps you design for production not for pse the the time off we're testing out it's it's a nice let's do a imagenet competition on on certain kind of devices is over let's do business then i can only say thank you very much julie and i hope that you that you will be able to convince many people that the time is now thank you thanks very much cheers thank you bye

Rise of AI 2027 · Berlin · May 2027

Be in the room for the next one.

Ten years of talks started as conversations in one curated room in Berlin. The next edition brings 300 selected leaders together in May 2027.