Quantifying Degradations of Convolutional Neural Networks in Space Environments

Advances in machine learning applications for image processing, natural language processing, and direct ingestion of radio frequency signals continue to accelerate. Less attention, however, has been paid to the resilience of these machine learning algorithms when implemented on real hardware and subjected to unintentional and/or malicious errors during execution, such as those occurring from space-based single event upsets (SEU). This paper presents a series of results quantifying the rate and level of performance degradation that occurs when convolutional neural nets (CNNs) are subjected to selected bit errors in single-precision number representations. This paper provides results that are conditioned upon ten different error case events to isolate the impacts showing that CNN performance can be gradually degraded or reduced to random guessing based on where errors arise. The degradations are then translated into expected operational lifetimes for each of four CNNs when deployed to space radiation environments. The discussion also provides a foundation for ongoing research that enhances the overall resilience of neural net architectures and implementations in space under both random and malicious error events, offering significant improvements over current implementations. Future work to extend these CNN resilience evaluations, conditioned upon architectural design elements and well-known error correction methods, is also introduced.