This repo contains API definitions to enable control and monitoring of AI powered complex software systems. It is part of a public funded research project named KI-Cockpit (AI cockpit) and more information can be found here https://www.kicockpit.eu/.
AI systems must ensure bias free data processing. Growing software system complexity makes it very hard even for expert observers, to understand decisions taken by modern software systems. Due to their stochastic nature AI algorithms add another layer of uncertainty. Despite the promising potential that these algorithms can add to the capabilities of software systems, measures needs to be taken, to keep humans in control.
Thus a two-fold API is proposed to guarantee human control. This API aims at being useful for any decision making system and hence is abstracting from individual – ever progressing – algorithms. The two function areas are centered around
1 Decision API: decisions taken in software systems
- Definition decision
- manual vs automatic decision
- necessary data
- Decision composition
- agnostic to decision method/algorithm
2 Transparency API: explain data processing and bias control
- Bias Check API
- pseudo randomized integration test results
- model info
- Algorithmic modules
- Descriptive Statistics
Requirements and design considerations are described here. Actual API definitions are in folder aicockpit-api - all necessary files for API definitions and code generation are provided.
The “KI-Cockpit” (AI Cockpit) project was funded by the Federal Ministry of Labor and Social Affairs and executed by the nexus Institute, Aalen University, the University of Stuttgart, Chemistree, Caritas Dortmund & Starwit Technologies with the support of Awesome Technologies Innovationslabor, the Institute for Innovation and Technology (iit) at VDI/VDE Innovation + Technik and keytec.
Software in this repository is licensed under the MIT license. See license agreement for more details.