The subtopic being addressed identifies current spacefaring computer hardware as insufficient for executing conventional artificial intelligence (AI) algorithms due to space, weight, and power constraints. Conversely, neuromorphic computing architectures have exhibited the ability to performatively execute AI programs while meeting these criteria. Presented here is one such general purpose neuromorphic computing architecture.
Based on the continuous time recurrent neural network model and instantiated upon the reconfigurable fabric of a field-programmable gate array, clusters of hardware-accelerated neurons can be evolved in real time while responding directly to environmental conditions. Preliminary work with this neuromorphic solution exceeded expectations when solving complex time-series problems while simultaneously minimizing spatial and power consumption.
Unlike many existing machine learning methods, this architecture can undergo hardware evolution for novel solutions or hardware adaptivity for existing solutions that are performing below necessary thresholds. Circuits undergoing intrinsic hardware evolution or adaptation exhibit naturally occurring fault tolerances as a result of real world environmental noise. These inherent phenomena make the continuous time recurrent neural network in evolvable hardware a powerful candidate for extraterrestrial and spacefaring operation.
NASA applications include: lightweight centralized sensory-affectory system for evolutionary robotics applications, energy-efficient neuromorphic implementation for cognitive radio networks, and a fault tolerant and adaptive onboard navigation system for planetary exploration.
Immediate commercial applications include: adaptive electromyographic interpreter for prosthetic limbs, energy efficient sensor preprocessor for internet-of-things (IoT) networks, and high performance analog radio frequency filter for cognitive radio.