Associative Memory in Spiking Neural Network Form Implemented on Neuromorphic Hardware

Abstract

Implementing cognitive algorithms on robots is one potential direction to realize autonomous artificial agents. As society and technology progress there is an effort to push robotics and artificial intelligence into many aspects of daily life. An important step in this process is leveraging concepts known to work from human cognitive features on computer systems. This paper shows an associative memory in the form of a spiking neural network (SNN), an application of the associative memory, and some performance benchmarking. SNNs allow these computational models to be instantiated in a low size, weight, and power (SWaP) form factor due to the biological efficiencies they approximate. The model is created in a neural network simulator and run on a low SWaP CPU and Intel’s Loihi, an artificial intelligence accelerator highly optimized for spiking neural algorithms. In addition, the model is employed on a mobile robotic platform that explores the real world and uses online learning to make associations. When the model was run on Loihi the overall power usage decreased as well as the run time of the simulation as compared to the low SWaP CPU proving beneficial to implement the neuromorphic hardware.