Deep neural networks are proven to yield state-of-the-art results on several complex computer vision tasks also relevant for autonomous driving. While error rates are promising, such models often require a large amount of computation at inference time, which poses a challenge for real-time applications or when hardware resources are limited on an embedded device. Within this topic we explore the problem of neural network pruning – methods to reduce complexity and inference time of deep models while keeping prediction quality at a level comparable to the original model.