Deep learning methods are used in many application areas and work very efficiently after a training phase. However, in general no reliable statement can be made about their correctness. In contrast, the correctness of analytical, plan-based or deductive algorithms can be verified with proven methods, but they are often too slow in application.
The project Fast&Slow investigates a systematic combination of both methods, in which the symbolic method trains the subsymbolic method and monitors its operation. The developed methodology will be validated in two extensive case studies in the areas of action planning in a smart home and trajectory planning for a two-arm robot.
The approach of combining the verifiability of symbolic approaches with the efficiency of sub-symbolic procedures can be applied in many application domains, since lack of verification is an acute obstacle to the use of deep learning techniques in safety-relevant areas such as autonomous driving.
