(De)centralized AI Solutions Technological and commercial design parameters Patrick Schmid Matthias Lederer ISM Munich, Germany ISM Munich, Germany Patrick.Schmid@ism.de Matthias.Lederer@ism.de ABSTRACT concept that has been introduced in 2017 (the Developments concerning artificial intelligence corresponding algorithm already in 2016) is to and machine learning have gained a lot of traction parallelize computing for machine learning in recently. Originally thought of as originating from order to decrease the latency [4]. a centralized device like a mainframe computer, a This development can be put into the context of major modification has been introduced in 2017 decentralized autonomous organizations [5] since with the “federated learning” concept performed it is comparable to the fundamental idea of cross- by decentralized agents. As this constitutes a functional working in a BPM (Business Process relatively new development, this research-in- Management) context. The often mentioned prime progress contribution addresses the topic of example is the parallelized machine learning for centralized vs. decentralized artificial intelligence image recognition on smart phone agents [4]. This in a broader context and in more detail. After a short paper discusses major topics of centralized brief introduction of the new concept with a vs. decentralized AI in more detail and displays mentioning of accompanying developments, a further application areas. selection of most relevant technological aspects as There are several accompanying developments well as commercial design parameters to decide on that encourage decentralization: for example (in particular the smart factory setting) is outlined. “softwarization” and “disappearing internet of things” [2]. The idea is roughly that devices are KEYWORDS more becoming like terminals as end-nodes of a Artificial Intelligence, Decentralized Autonomous Organization, network, thus spanning a seamless continuum Disappearing Internet of Things, Embedded AI Solutions, Federated Learning, Machine Learning, Smart Factory, from the devices to the cloud [2]. Here in Softwarization. particular a lot of computation already takes place at the devices, so that computation is moving 1 Introduction and Background “toward the edge”, hence enforcing Edge and Fog Computing (here Fog Computing covers the gap As artificial intelligence (AI) applications between Cloud and Edge Computing) [2]. As typically involve machine learning (ML) routines simple computations can typically evolve into relying on the processing of big data, the time machine learning routines, there seems to be a needed for computational tasks is seen as a major similar development for decentralized AI constraint [1]. This applies even more if the solutions in a broader context (AI at the edge learning procedures are performed on a central instead of simple computations at the edge) and device like a mainframe computer [2], [3]. One of not only in the aforementioned image recognition the main ideas behind the “federated learning” example. So the decentralized devices can not only S-BPM ONE 2019, June 26-28, Sevilla, Spain Patrick Schmid, Matthias Lederer be seen as contributors to a consolidated learning considerations, federated learning can procedure, but in fact learning for their own constitute one method to avoid bottlenecks and purpose as well as exchanging information and/or speed up critical process steps. synchronizing when it seems useful. In short, AI is 2. Quality: Apart from “just” parallelizing tasks no longer purely centralized and also no longer for ML procedures, where the volume of the data that has to be processed is very large, the purely a result of one single participant in learning. learning of several agents can be used to So, there are two distinct conceptual setups: either improve the overall quality of the respective several agents contribute to a consolidated ML routine [6]. Typically, this is done via learning in a hierarchical structure without ensemble methods of machine learning (for substantially learning for themselves (for example instance via “bagging” or “boosting” to name when they only perform a limited number of a few examples) [7]. One might question iterations [6]) or several agents synchronize on the whether quality is strictly improved by same hierarchy level meaning that they produce an consolidating the inputs of various learning aggregated model that is shared between all of agents (as they might provide misleading them. adjustments). The quality improvement cannot be strictly guaranteed and is dependent on the method used for aggregation, but typically, 2 Relevant technological aspects more input from agents should generate the There are several relevant (technological) aspects potential for superior quality in learnings [7]. with accompanying challenges that have to be Bagging and boosting may be the most widely known procedures for this, but of course there understood and mastered in order to decide on the exist many different variants with their own parameters of central/decentral AI approaches in a nuances / strengths and this topic is very much meaningful manner: an area of active research [7]. Detailed analysis is needed what specific algorithmic procedure 1. Speed: As already mentioned, multiple agents in what setup best leverages the already gained can be regarded as a kind of parallelized insights of the single distributed learners (to computing mechanism in order to use illustrate the nuances of such a procedure, it available time windows more efficiently. This shall be mentioned that the newly developed objective is in particular achieved via federated learning concept also considers delegating subtasks in a hierarchical setup to which subparts of the training set are sent to various agents hence improving “virtually” the the distributed clients) [4]. computing power. Furthermore, as the 3. Compatibility: The agents that participate in computational procedures are performed at the hierarchical decentralized learning or same local clients at the edge (in the sense of Edge level synchronization must adhere to the same Computing), several data transmissions and machine learning method / model as a information exchanges with the cloud become necessary condition in order to facilitate obsolete [6]. In summary, speed is not only consolidation and exchange of information increased by higher computational power due (for example weight functions). Of course, as to the number of participating devices, but also there exists a huge variety of machine learning by reduction of now unnecessary back-forth procedures which are all performed transmissions of the training data which is differently, these methods need to be aligned. stored locally [6]. Additionally, a reduction of The same applies to questions of (where applicable) communication costs can implementation, since, for example, in a smart occur [6]. So under business process factory various proprietary devices with (De)centralized AI solutions S-BPM ONE 2019, June 26-28, Sevilla, Spain underlying architectures and software [9] in science and business practice: the solutions have to work together [8]. “embedded AI solutions” and applications in the 4. Security/Privacy: The security challenges smart factory setting. arise when a lot of not fully trustworthy agents contribute to the decentralized learning (either 3.1 Embedded AI Solutions hierarchical or same level) [6]. One can refer There is one related thinking approach by C. E. here to discussions about security in service- Bouee [10] that deserves special attention as he oriented process environments - there are was the first to address the decentralization issue similar challenges (for example, asking for a of AI in full force (in particular with respect to its trustful registry). It could be that information commercial dimension). He envisions an enforced supplied for consolidation is "manipulated" intentionally to distort the quality of learning development and usage of “portable AI” solutions [6]. Here lies the corresponding challenge of which are decentralized AI solutions that can be suppressing such occurrences. Apart from that integrated (embedded) into all kinds of devices, challenge, decentralized learning is commonly typically resembling a highly individualized smart regarded as a way to increase privacy since assistant [10]. Furthermore, he calls into doubt that confidential data can be kept locally and only the large digital platforms that dominate the tech processed (anonymized) information, like sector today will dominate the market for portable weightings of ML routines, is then shared [6]. AI solutions in the future and even believes that portable AI could end their monopoly [10]. Considering these technological aspects, it can be discerned that there are several potential He provides several examples for such emerging advantages which can be gained from the companies that are active in the segment of application of decentralized learning scenarios. embedded AI solutions, thereby illustrating the However, also respective challenges need to be parameter decision of (de)centralization [10]: addressed in order to fully realize these potential advantages. 1. Snips (from France) attempts to put “AI in every device”, hence creating “a unified voice strategy with a complete range of interface offerings - from simple voice commands to 3 Commercial application areas and comprehensive natural language voice use cases recognition” [11]. So they are exactly addressing the softwarization topic already After understanding technological aspects, described. business questions and typical use cases that build on the decentralized learning idea shall be 2. Arago (from Germany) focuses on the B2B addressed. Generally, decentralized artificial sector and its solutions are meant “to automate intelligence can be applied to any field where enterprise IT and business operations” via currently artificial intelligence solutions are machines “with human problem-solving deployed since there are typically multiple devices skills” that underwent supervised training in use which are up to now not connected regularly [12]. Hence, they essentially support the in a hierarchical or same level relationship. In the automation of business processes and show the potential of decentralized AI solutions therein. following, two selected areas are outlined since they play a major role in the ongoing discussion S-BPM ONE 2019, June 26-28, Sevilla, Spain Patrick Schmid, Matthias Lederer 3. SenseTime (from China) is active in many Highly individualized smart assistants are then AI deep learning and supercomputing areas, in solutions in their most decentralized form as these particular with its independently developed solutions do not just complement technical large-scale training system it also employs devices, but also every single end user [10]. Bouee decentralized machine learning [13]. elaborates that basically new business models will Therefore, they act as a kind of orchestrator for arise from this as these individualized smart decentralized learning. assistants can act as much more strict gatekeepers From this exemplary listing one can already get an for personal decisions than common smart idea in what fields entrant companies try to capture assistants currently do [10]. An interesting market share from the incumbent large tech immediately arising question is, whether companies. Foremost, they are supposed to have a companies can and will deliberately set up barriers competitive advantage over the incumbent such that personalized smart assistants cannot be companies since they are not burdened with used to full extent (much like the already existing having negative publicity concerning data privacy bot-barriers for travel websites). and security issues. Furthermore, it is conceivable that they gain advantages from more appealing The topic of decentralization is of particular product offerings with higher individualization relevance in a likely “second wave” of and “tailor-made” fit [10]. One further argument digitization. While in the first (B2C marketing- of Bouee for the competitive edge of these oriented) wave large (U.S.) companies use emerging firms is that data becomes obsolete centrally stored personalized data, manufacturing faster and that therefore the advantages which processes and B2B transactions will then be incumbent companies possess from their already digitized in a second wave. Here, the tendencies to amassed data consequently diminishes very centralize AI will likely not take place due to quickly [10]. However, on the contrary it can be already established physical equipment (with argued that the large incumbent companies with various therein embedded software standards) their already existing wide outreach and which is difficult to merge. Furthermore, the computational power do still retain an advantage possible fear of process participants of losing in this respect since bigger amounts of data are control over their data encourages created faster and hence require much more decentralization. computational power for assessing and evaluating 3.2 Smart Factory these large volumes. This is certainly the case with The other large application field that should be respect to potent cloud offerings which heavily discussed in more detail is the smart factory setting rely on the magnitude of computational strength. which was originally envisioned by the German Ministry of Education and Research (BMBF) [15]. The aspect of highly individualized smart “Smart factory” refers to manufacturing processes assistants is also very interesting, as this is (including resources) in which relevant involved somehow the opposite of a smart assistant machines and devices are all equipped with provided by a large technology company that sensors and perform typical smart tasks, like could be sometimes prone “to lose its voice” (as Condition Monitoring, Diagnosis for Maintenance featured in a famous commercial) [14]. and Optimization of Processes, that can be improved via machine learning [16]. It is certainly (De)centralized AI solutions S-BPM ONE 2019, June 26-28, Sevilla, Spain conceivable that an enterprise has multiple smart concern certain types of machines, whereas factories in use that either share the full Optimization of Processes might depend on the infrastructure or at least key components. This is level in a multi-level production process. then the “factory network” scenario that strives to However, if a sufficient similarity can be optimize network performance [17]. This typically guaranteed, such (de)centralized learning concepts addresses the four basic layers of the “Life Cycle make a lot of sense. One could argue that different Value Stream” of the RAMI 4.0 framework: tasks are never similar enough, however, Asset, Integration, Communication and considering that a factory is a highly standardized Information [18]. setting, it is conceivable that there might be a As data basically is the “lifeblood” of smart smart transformation/translation of corresponding factories, an evident approach for maximizing use cases. network performance would be to harness the available data of the whole network in a more But not only the question “whether” it does make efficient way by centralizing the decentralized sense, also the question “how often or regular” learnings of specific similar tasks. such a synchronization shall take place has to be answered. In a mimimum scenario, the Here one has to carefully distinguish according to synchronization can take place in typical inactive the degree of similarity. To give two extreme periods like production pauses or overnight as not examples: Concerning autonomous driving it to interfere with the regular business operations makes a lot of sense to align and share learnings, (thereby reducing process speed). This is in since increasing the security of autonomous particular apt for devices and components used in driving has highest priority and conditions of process steps that require fast decisions and streets are highly standardized. However, if actions. However, if devices and components are natural language processing is concerned, it seems operating on a much slower time scale (for recommendable not to mix and interchange example registering inbound commodity flows or training data or learning weights of different outbound logistics of finished products) the dialects within a language (although they do synchronization could take place continuously. represent the same language) as this rather tends Similar questions are relevant when integrating to create confusion instead of resolving it. partners in SCM (Supply Chain Management) or In the same manner one has to proceed in the smart PLM (Product Lifecycle Management) processes. factory setting, since there might be tasks and However, in modern value networks which rely on routines that are more similar to each other than to data exchange for smart/reliable decisions, well- the rest (depending on the type of smart factory). known process standards (e.g., EDI, flat XML) Certainly, such training and synchronization may not be sufficient. procedures can be applied to the same type of task within the same type of factory. But generalization According to experts of “Internet of Business” the to similar tasks is, as always, accompanied by 10 smart factory trends to watch in 2019 are the certain challenges in deciding whether the following [19]: similarity is sufficient enough. Here it is conceivable that differences for Condition Monitoring and Diagnosis for Maintenance S-BPM ONE 2019, June 26-28, Sevilla, Spain Patrick Schmid, Matthias Lederer 1. “Collaborative robots will augment agendas. The strength of such shared learnings workforces” should in particular become apparent in cases 2. “Cloud robotics & APIs will give where there is a lack of suitable training material, manufacturers greater control” for example when whole processes in smart 3. “Robotics-as-a-service will make robotics factory are redesigned according to machine viable for smaller manufacturers” learning insights. Although this would be a 4. “5G & Multi-access Edge Computing (MEC) will help keep factory workers informed” desirable feature of the smart factory, this redesign 5. “Edge computing will see new use cases” of processes is certainly regarded as a much more 6. “Cybersecurity will be given greater priority” complex task than just checking the need for 7. “AI & advanced analytics will become near- maintenance of a machine since much more ubiquitous” factors (like implications for following processes 8. “Digital twins will be employed more widely and business partners) have to be considered and across manufacturing and the supply chain” evaluated. 9. “Additive manufacturing will be used to create final products” 10. “Wearables will become commonplace on the 4 Conclusion factory floor” In this short paper relevant design aspects and implications of connecting formerly unconnected In this list there are two trends that have a AI solutions with each other were discussed as particular high relevance for our discussed topic, well as the topic of creating much more agents that namely “Edge computing will see new use cases” are amenable to such setups. The connection of AI and “AI & advanced analytics will become near- solutions can either be understood as delegating ubiquitous”. By design, the components and tasks in a hierarchical setup or as a kind of devices of a smart factory are meant to make synchronization between agents on an equal autonomous decisions. So in time-critical footing. The potential of such setups that processes it is wanted that more computation is exemplify a much greater extent of moving towards the edge in order to increase the interconnection has yet to be realized, but it is speed and decrease the latency. Also crucial is the expected to be adopted and enforced much more rising importance and spread of AI & advanced in upcoming developments, for example on the analytics, since this first ensures that local AI way to the fully integrated smart factory. solutions are present and, furthermore, that there exists potential which can be realized via their alignment. It is certainly highly dependent on the particular use case whether the cooperation of the distributed AI solutions takes place on the same level or in a “master-slave” setup. But it has to be noted that the aforementioned centralization procedures for AI solutions are not yet to be anticipated in full force for the current year, hence pointing out potentials for realizations that still have to find their place on the respective (De)centralized AI solutions S-BPM ONE 2019, June 26-28, Sevilla, Spain REFERENCES from https://www.bmbf.de/de/zukunftsprojekt- [1] E. Alpaydin. 2014. Introduction to Machine Learning industrie-4-0-848.html. (3rd. ed.). The MIT Press, Cambridge, Chapter 1. [16] A. Maier, O. Niggemann and S. Schriegel. 2017. [2] A. Manzalini. 2015. Softwarization and the Big Data and Machine Learning for the Smart Factory— Disappearing Internet of Things, Retrieved from Solutions for Condition Monitoring, Diagnosis and https://iot.ieee.org/images/files/pdf/softwarization_and_ Optimization. Industrial Internet of Things, 473–485. iot DOI: https://doi.org/ 10.1007/978-3-319-42559-7 _webinar_24022015.pdf. [17] R. Burke, M. Hartigan, S. Laaper, A. Mussomeli [3] C. Wurster. 2002. Computers: An Illustrated History. and B. Sniderman. 2017. The smart factory (Deloitte Taschen, Cologne. Insights). Retrieved from https://www2.deloitte.com/ [4] Company announcement. 2017. Retrieved from insights/us/en/focus/ industry-4-0/smart-factory- https://ai.googleblog.com/ 2017/04/federated-learning- connected-manufacturing.html. collaborative.html. [18] ZVEI (Zentralverband Elektrotechnik- und [5] A. Kiulian. 2017. Why Decentralized Artificial Elektronikindustrie e.V.). 2015. Das Intelligence Will Reinvent The Industry As We Know It. Referenzarchitekturmodell RAMI 4.0 und die Industrie Retrieved from https://www.forbes.com/sites/ 4.0-Komponente. Retrieved from forbestechcouncil/2017/11/16/why-decentralized- https://www.zvei.org/themen/industrie-40/das- artificial-intelligence-will-reinvent-the-industry-as-we- referenzarchitekturmodell-rami-40-und-die-industrie- know-it/#74f6f300511a. 40-komponente/. [6] Mike (@mikepqr). 2018. Federated learning: distributed [19] A. Hobbs. 2018. Complete guide: 10 smart factory machine learning with data locality and privacy. trends to watch in 2019 (Internet of Business), Retrieved Retrieved from https://blog.fastforwardlabs.com/ from: https://internetofbusiness.com/complete-guide- 2018/11/14/federated-learning.html. 10-smart-factory-trends- to-watch-in-2019/. [7] Z. Zhou. 2012. Ensemble Methods: Foundations and Algorithms (1st. ed.). Chapman and Hall/CRC, New York, Chapters 2,3. [8] A. Abdi, M.R. Abdi, F.D Edalat and A.W. Labib. 2018. Integrated Reconfigurable Manufacturing Systems and Smart Value Chain. Springer, Cham, Chapter 4. [9] S. Betz, M. Kurz, M. Lederer and W. Schmidt. 2017. Some say Digitalization - others say IT-enabled Process Management thought through to the End. Proceedings of the S-BPM ONE '17 (C. Zehbold and M. Mühlhäuser (Ed.)). ACM Press, New York, NY, USA. DOI: https://doi.org/10.1145 /3040565.3040574 [10] C.-E.Bouee. 2018. Smarter than Man Friday (Think:Act Magazine #24). Retrieved from https://www.rolandberger.com/en/Insights/Global- Topics/ Artificial-Intelligence/ [11] Snips company website. 2019. https://snips.ai/. [12] Arago company website. 2019. https://arago.co/arago/. [13] SenseTime company website. 2019. https://www.sensetime.com/. [14] C. Gartenberg. 2018. Amazon has a clever trick to make sure your Echo doesn’t activate during its Alexa Super Bowl ad. Retrieved from https://www.theverge.com/2018/2/2/16965484/amazon- alexa-super-bowl-ad-activate-frequency-commercial- echo. [15] German Ministry of Education and Research (BMBF). 2013. Zukunftsprojekt Industrie 4.0. Retrieved