Fraunhofer Institute for Cognitive Systems IKSArtificial Intelligence (AI) is now ubiquitous in the technology landscape. In some engineering application domains, such as manufacturing, it is very important that AI techniques produce accurate results. Trustworthy and reliable performance means that AI can be used for high-level decision making and in safety-critical scenarios.
© iStock.com/zorazhuang
Trustworthy and Dependable AI Applications
Challenge: Complexity and Uncertainty
- Ensuring reliability in AI applications is challenging for two main reasons: The complexity of the techniques and the amount of uncertainty they introduce.
- Artificial Intelligence techniques are complex due to the sophisticated algorithms they implement, the parameters and architectures they use, and the general difficulty in understanding and interpreting the decisions they make.
- At the same time, these AI techniques rely on different data sets, stochastic search heuristics, and training and operational contexts, all of which introduce multiple sources of uncertainty.
Need: Principles Applicable to AI
- The key to overcoming these challenges is to identify a set of principles that can be applied to increase the reliability and trustworthiness of AI techniques.
- These principles should be generalizable enough to be applied to multiple techniques but should also provide concrete validation that the AI meets the desired criteria.
- If they can strike this balance, they can be applied broadly to increase reliance of AI applications in business.
Solution: Adapted Techniques
A pragmatic approach to ensuring trustworthiness and reliability in AI applications is to identify applicable principles from other domains and adapt them to this context. For example, safety standards can be developed specifically for AI. At the same time, work to improve the interpretability and understandability of AI will allow users to follow the decision-making process behind the techniques, thereby increasing confidence in the results produced.