AutoDeep: Automatic Design of Safe, High-Performance and Compact Deep Learning Models for Autonomous Vehicles

Purpose and goal

Deep Neural Networks (DNN) are increasingly being used to support decision-making in autonomous vehicles. High DNN accuracy comes at high computations, storage, and memory bandwidth requirements, which makes their deployment particularly challenging, especially for vehicle embedded computing platforms. In this project, we will develop an automatic framework to achieve performance, compactness, and robustness in design and customization of DNN for safety-critical applications such as intention detection for road and construction autonomous vehicles.

Expected effects and result

The project expected results are design and optimization methodologies for DNNs in safety-critical vehicular applications. The project will provide new knowledge to extend the state of the art in the area of deep learning applications. The results are also expected to have a considerable impact on the state of the practice with regards to new and extended development models, prototypes and industrial demonstrators. The plan is to disseminate the results also to broader audience in the form of international publications. The project is expected to produce two Licentiate theses.

Planned approach and implementation

The first work package (WP) is on design and optimization of DNNs models for intention detecting. To choose the embedded friendly architecture, we will develop an automatic framework, which designs a highly optimized set of DNN architectures to be deployed on embedded computing platforms. The second WP is on safety and robustness of DNN models where we consider the effect of the adversarial and unintended data perturbations on the robustness of DNNs. The third WP is aimed at evaluating AutoDeep using the in-house testbed of an autonomous vehicle along with industrial case studies.