ConformalHDC: Uncertainty-Aware Hyperdimensional Computing with Application to Neural Decoding

A research team from UC Irvine and UC San Diego has introduced ConformalHDC, a novel framework that integrates conformal prediction with Hyperdimensional Computing (HDC) to address a critical limitation in neuromorphic systems. While HDC offers computational efficiency for brain-inspired computing, it traditionally lacks rigorous uncertainty quantification, leaving systems vulnerable to outliers and adversarial attacks. ConformalHDC solves this by providing statistical guarantees through two complementary approaches: a set-valued formulation that creates enclosed decision boundaries with finite-sample coverage guarantees, and a point-valued formulation that can improve accuracy by accounting for class interactions.

The team demonstrated ConformalHDC's effectiveness on the challenging task of decoding non-spatial stimulus information from hippocampal neuron recordings during sequence memory tasks. The system not only accurately decoded neural activity but also provided reliable uncertainty estimates and correctly abstained when presented with data from different behavioral states. This breakthrough enables more robust brain-computer interfaces and neuromorphic hardware that can operate reliably in real-world conditions where uncertainty is inherent. The framework's computational efficiency combined with statistical rigor makes it particularly suitable for edge devices and implantable neural interfaces that require both accuracy and safety guarantees.