ProjectSMILE II - Safety analysis and verification/validation of MachIne LEarning based systems

SMILE II - Safety analysis and verification/validation of MachIne LEarning based systems

SMILE II - Safety analysis and verification/validation of MachIne LEarning based systems

To increase the level of autonomy, future vehicles will depend on functions that rely on Deep machine learning (DML) whose correct behavior cannot be guaranteed by traditional software engineering approaches. Furthermore, crucial parts in ISO 26262 are not well defined for addressing autonomous systems, and certain process requirements and recommendations are not applicable for the development of machine learning in the domains of specification, design, and testing. Based on the findings from the SMILE concept study (Englund et al. 2017), we have identified a series of opportunities that can improve the robustness when utilizing DML-based perception for autonomous vehicles.

The SMILE II project focuses on developing methods that allow DML-based functions to be included into safety critical vehicular applications. The proposed approach in this application constitutes the second out of three planned parts of the research endeavour, i.e., a three stage SMILE research program. In this project (SMILE II), we focus on designing a safety cage encapsulating the DML component. The focus in this project will be around verifying that the input data that is captured during run-time (from perception sensors) is representative for the model, thus, the models should only act on data that comes from the same distribution as that used for training the models. Thus, during training, the data that goes into the DML is something that it recognises and is trained to act upon. The safety cage will increase the reliability of the predictions by ensuring that the DML acts within a valid design space. In addition, we aim to explore how the data from outside of the training region of the models, could be used for updating models.

The next phase of the SMILE research program (SMILE III) will concern the actions to be made if the data is considered to come from outside of the training space, i.e., the design of an alternative model for handling this data should be made. Additionally, the aim is to have (at least) 3 PhD students associated to the program, enrolled at e.g. Chalmers, Lund University and Halmstad University. The results from the project are of interest for actors that want to use machine learning in safety critical applications. 

The main applicant of this project is RISE Viktoria, and other partners are RISE SICS, QRTECH, Semcon, Volvo Cars and Volvo Technology AB. The project will run from October 2017 to September 2019 and is funded by Vinnova within the FFI - Machine Learning call.