Influential Learning : Knowledge Sharing between Artificial Neural Networks for Autonomous Vehicles

Abstract: Autonomous vehicles may be a part of our future no matter if we like it or not. The technology developed for self-driving have already outperformed humans in multiple aspects but involves systems that are prone to failure. Machine learning techniques have proven to enhance the performance of these autonomous vehicles by efficiently analyzing and learning from data gathered through embedded systems. A popular approach to enhance performance in autonomous systems is to incorporate cameras and use computer vision techniques to extract useful information from the images. The potential use cases of applying such techniques are many including object detection, tracking, segmentation, motion estimation and scene understanding. Numerous implementation methods have been proposed with the ambition to optimize these techniques in order to utilize their full potential and help autonomous systems maneuver as accurate as possible. Velocity and steering angle are important aspects of autonomous vehicles, however since the vehicles can reach high velocities it is also of great importance that the systems act quickly in addition to being accurate. This thesis evaluates an approach that involves combining machine learning models in a way that enables one of the models to be influenced by the others with knowledge unobtainable by itself. While learning to estimate a steering angle one of the models is simultaneously taught to mimic auxiliary models that have achieved state-of-the-art performance in tasks related to image segmentation and optical flow. The machine learning models used are convolutional neural networks and the intention is for one of the neural networks to acquire the knowledge of the optimal direction to steer a vehicle. The project conducted in this thesis shows that the performance on gathered test data can be significantly improved with the proposed approach while all used neural networks are able to handle the data. After training has been conducted the auxiliary neural networks can be discarded and therefore this approach achieves the same processing time and memory size as if they were not involved.