IoT Tech Expo 2023_Hans-Martin Heyn presentation
Download as PPTX, PDF0 likes16 views
The document presents a compositional architecture framework for distributed AI systems, emphasizing co-design and collaboration among various stakeholders. It outlines challenges in developing AI systems within an agile framework, focusing on the need for a structured approach that incorporates different design aspects and non-functional requirements. The framework aims to ensure safety, security, and fairness while facilitating traceability and compliance with emerging AI regulations.
IoT Tech Expo 2023_Hans-Martin Heyn presentation
- 1. Hans-Martin Heyn University of Gothenburg, Sweden Requirements Analysis and Decomposition for Distributed Systems based on Deep Learning A compositional architecture approach 26. September 2023 Eric Knauss, Hans-Martin Heyn, Olof Eriksson, Oliver Brunnegard, Stefan Andersson
- 2. 2 Context
- 3. 3 Context
- 4. 4 Concerns in (AI) system design • AI, and especially highly complex DNN, cause more concerns to be included in the system design. • A reason for the additional concerns can be uncertainty in how to build these systems right.
- 5. 5 • Co-design can stand for “collaborative design”. • All required stakeholders should be actively involved in the design process. – Developers of different disciplines, Data Experts, Customers, Business Owners, … • See for example – Fitzgerald, J., Larsen, P.G., Verhoef, M., 2014. Collaborative design for embedded systems. Academic Press 10, 978-3. – Nalchigar, S., Yu, E., Keshavjee, K., 2021. Modeling machine learning requirements from three perspectives: a case report from the healthcare domain. Requirements Engineering 26, 237-254. undraw.co Co-design of a system
- 6. 6 • Co-design can mean “integrated design”. • Different design aspects of the system, e.g., hardware, software, but also quality aspects are closely coupled to each other. • This can create a high dimensional design space that needs to involve in parallel to ensure “safe by design”, “secure by design”, “fair by design”, or any other “quality by design” necessary. undraw.co Co-design of a system
- 7. 7 • Designing a complex and distributed system is a hierarchical process. – Several, sometimes highly specialized views allow for decomposition of the design task. – Requirements and architecture often co- evolve (Twin Peaks). • Developing complex system is a highly collaborative act between many stakeholders. Co-design of a system undraw.co
- 8. 8 • We needed to define an architectural framework, that supports both aspects of co-design. • The framework must support explicitly aspects of distributed systems (IoT) and AI system development. – Learning and data management • The framework also needed to be flexible enough to cover all current use cases, and new future use cases. – A special focus therefore lies on the support of non- functional requirements / quality views – Traceability of design decisions • A single reference architecture would have been too limiting for allowing the variety of (open) use cases in VEDLIoT. Co-Design Integrated design Collaborative design Problem Definition
- 9. 9 Require- ments & Context Data Ingestion Data Prep- aration Training Validation and Testing Deploy- ment & Monitoring • Developing AI systems in an truely agile software project is challenging • Strong dependencies exist between different steps in a typical ”AI/ML development pipeline”. AI Dev. Pipelines and Agile?
- 10. 10 Require- ments & Context Data Ingestion Data Prep- aration Training Validation and Testing Deploy- ment & Monitoring AI Dev. Pipelines and Agile?
- 11. 11 Require- ments & Context Data Ingestion Data Prep- aration Training Validation and Testing Deploy- ment & Monitoring AI Dev. Pipelines and Agile?
- 12. 12 Require- ments & Context Data Ingestion Data Prep- aration Training Validation and Testing Deploy- ment & Monitorin g Solution Space Problem Space AI Dev. Pipelines and Agile?
- 13. 13 Solution Space Problem Space Requirements & Context Data Ingestion Data Prep- aration Training Validation and Testing Deployment & Monitoring Validation and Testing Validation and Testing Compositional thinking for architecture frameworks
- 14. 14 Solution Space Problem Space Compositional thinking for architecture frameworks
- 15. 15 Solution Space Problem Space Compositional thinking for architecture frameworks
- 16. 16 Knowledge creation (e.g. definition of security goals). Concept design (e.g. introduction of safety mechanisms). Final design (e.g. assigning functions to secure processor environments. Monitoring concept definition (e.g. monitoring of secure processing environment). Level of abstraction - Example
- 17. 17 Rule 1: Clusters of concern shall contain architectural views with different levels of details of a certain aspect of the system under development. Rule 2: Architectural views shall be sorted into levels of abstractions, according to their level of details about the system under development. Rule 3: By using correspondence rules, it shall be possible to arrive at different architectural views of the system without encountering inconsistencies. Rule 4: Architectural views, and relations between them, shall be mapped to the next lower level of abstraction. Compositional thinking for architecture frameworks
- 18. 18 Iterative and middle-out design of AI systems • Step 1: Identify clusters of concern • Step 2: Identify levels of abstraction • Step 3: Add existing architectural decisions. • Step 4: Add missing architectural views. • Step 5: Add missing relations. • Step 6: Iterate if needed.
- 19. 19 Step 1: Identify clusters of concern Step 2: Identify levels of abstraction Step 3: Add existing architectural decisions Step 4: Add missing architectural views Step 5: Add missing relations . . . ... Step 5: Add missing relations Traceability
- 21. 21 A compositional architecture framework for VEDLIoT
- 22. 22 • The architectural framework helps connecting different aspects of a (distributed) AI system together. • It is based on mathematical ideas from category theory and compositional thinking. • It allows for “middle-out” development, i.e., existing design decisions are explicitly considered. • It allows to keep an overview over the necessary quality aspects, such as safety, security, fairness, or privacy aspects of the systems. • The framework enforces a runtime concept for the system. • The traceability of design decisions allows for compliance with upcoming AI regulations. Summary
- 23. 23 Thank you for your attention.