Efficient Implementation of Spiking Neural Networks on the Loihi Neuromorphic Processor for Autonomous Driving Problems

Autonomous Driving (AD) related features provide new forms of mobility that are also beneficial for other kind of intelligent and autonomous systems like robots, smart transportation, and smart industries. For these applications, the decisions need to be made fast and in real-time. Moreover, in the quest for electric mobility, this task must follow low power policy, without affecting much the autonomy of the mean of transport or the robot. These two challenges can be tackled using the emerging Spiking Neural Networks (SNNs). When deployed on a specialized neuromorphic hardware, SNNs can achieve high performance with low latency and low power consumption. In this thesis we face two important problems in AD field i.e. the classification between cars and other objects and detection of the street lanes. For these problems we use an SNN connected to an event-based camera. To consume less power than traditional frame-based cameras, we use a Dynamic Vision Sensor (DVS). For the classification problem we develop an accumulation strategy that is used to compress the incoming spiking information, in order to further reduce the latency and power consumption. The experiments are made following an offline supervised learning rule, followed by mapping the learnt SNN model on the Intel Loihi Neuromorphic Research Chip. Our best experiment achieves an accuracy on offline implementation of 86%, that drops to 83% when it is ported onto the Loihi Chip. The Neuromorphic Hardware implementation has maximum 0.72 ms of latency for every sample, and consumes only 310 mW. To the best of our knowledge, this work is the first implementation of an event-based car classifier on a Neuromorphic Chip. For the lane detection we made the experiments using also an offline supervised learning rule, followed by mapping the learnt SNN model onto the Intel Loihi Neuromorphic Research Chip. During the training for the loss function we use a new method based on the linear composition of Weighted binary Cross Entropy (WCE) and Mean squared error (MSE) measures. With that we can exploit the best for the supervised learning rule. We find good results with a maximum Intersection over Union (IoU) measure of about 0,62 and a very low power consumption of about 1 W. The developed system has also little complexity, in fact the best IoU is achieved with the implementation of only 36 neurocores. The Latency is low and the highest time to recognize an image is of less than 8 ms.