Latency Mitigation in Helmet Mounted Displays using Next Generation Neural Network Architectures

Neuromorphic Processing Units are non-Von Neumann machines differentiating them from other processing architectures.  This difference complicates traditional algorithm mapping as common methodologies are optimized for general purpose computing architectures.  The performance a given algorithm achieves can be greatly impacted by implementation and mapping onto the processing architecture.   The primary objective of this research effort is to determine the optimal methods for mapping algorithms to this architecture class.  Specifically, we will experiment with multiple neural network types and provide a quantitative characterization of utilizing Neuromorphic Processing Unit based processors to implement existing algorithms in the form of these types of neural networks.  Furthermore, one must understand the performance achievable on more common technologies to adequately characterize the pros/cons of the NPU. To that end, we also characterize more traditional computations platforms.  To that end, this work targets implementation and evaluation of specific items that we see as essential elements of a successful evaluation of NPU architectures.  Using existing RPA designed neural algorithms, we will utilize the Nengo framework to model NPU architectures and evaluate performance of multiple neural network representations of this algorithm.   Specifically, we will implement a fully connected Deep Neural Network (DNN), a recurrent neural network (RNN), a long short-term memory (LSTM), and a gated recurrent unit (GRU).   We will utilize Tensorflow for implementation of these networks to allow rapid evaluation of  competing architectures including CPU, GPU and TPU.  Additionally, we will develop benchmark and analysis software to provide performance metrics and reporting.