SafeHDC: Concurrent Uncertainty and Fault Detection in Hyperdimensional Computing

Abstract

HyperDimensional Computing (HDC) is a brain-inspired machine learning paradigm that encodes data into high-dimensional hypervectors and performs classification through similarity comparisons. Its computational efficiency, one-shot learning capability, and robustness to noise make it attractive for resource-constrained edge applications. However, HDC systems can still exhibit uncertainty when inputs deviate from the training distribution, such as with noisy or Out-of-Distribution (OOD) data, and when hardware faults occur within the encoding or classification stages, potentially compromising reliability in safety-critical scenarios. This work introduces SafeHDC, an ultra-lightweight, concurrent runtime mechanism that detects both input-induced uncertainty and hardware-induced anomalies without model retraining or added latency. SafeHDC identifies noisy and OOD inputs as well as internal faults with negligible hardware cost (2 LUTs, 1 register). Experimental results show that SafeHDC achieves competitive detection accuracy and reaches up to 100% fault-detection coverage in many situations, enabling real-time safety enhancement for HDC accelerators with minimal overhead.