A complete safety argumentation can be made via methods that support humans in understanding and analysing why an AI system took a specific decision. Moreover, insights into the inner operations of an AI system (DNN in particular) can increase the trust DNN in general; in fact, the human understanding about semantics helps in a semantic understanding of the DNN, and it is an iterative process.
Main Question
Have the relevant dimensions been founded and selected, which affect and influence the performances of the AI system? (Burton et al., 2022)
Sub-Questions:
- Has the use of a specific tool been considered? (Haedecke et al., 2022)
- Can this tool support and enable the user’s visual analyses over huge amounts of data in such a way to understand the nature of the underlying data and the DNN performance?
- Based on that, can the user make hypotheses about the relevance of semantic and non-semantic dimensions, which then can be checked or refined interactively?
- Will this result in human advantages in terms of major understanding and deeper knowledge of AI system behaviour?
References
- Burton, S., Hellert, C., Hüger, F., Mock, M. and Rohatschek, A. (2022) ‘Safety Assurance of Machine Learning for Perception Functions’. Deep Neural Networks and Data for Automated Driving, pp.335–358. doi:https://doi.org/10.1007/978-3-031-01233-4_12.
- Haedecke, E., Mock, M., S. Houben, M. Akila (2022) ScrutinAI: Visual Analytics Approach for the Semantic Analysis of Deep Neural Network Predictions. EuroVis Workshop on Visual Analytics. Available at: https://publica-rest.fraunhofer.de/server/api/core/bitstreams/f7d5aca7-5080-4153-b1e1-aa113032a4f3/content (Accessed: 27 May 2024)