Zhiliang Zhou
CEO
Al Metrics

Unlike traditional function development, many function development for the autonomous driving is learning-based and data-driven, which caused great challenges to both development and Validation process. For example, when using a neural network based model, it is difficult to say if the current training dataset is various enough to ensure the model with good generalization ability. Besides, many machine learning based models are improved iteratively, it is not possible to achieve Validation by letting a real car have test runs for every iteration

To support such development and Validation processes, we designed a learning support data platform, which improves the efficiency of  machine learning based development and Validation for perception and planning subsystems, this data platform consists of a data warehouse, algorithm modules, and corresponding APIs. The data warehouse uses a Database server as Label/attribute organizer and uses scalable file systems to store multimedia files ROSbag files etc. The algorithm modules are used to analyze model’s performance and detect model’s flaws.

For the development process, we propose a solution, which allow the developer to more targeted improve and tuning model, with help from organized data labels /attributes. Besides through algorithm module, we can make machine learning models more robust, e.g. improve the robustness of neural network based mode by adversarial example analysis.

To improve the Validation process, we propose a solution, which uses data playback simulator with reference models to validate given module or subsystem.