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.