Fault robustness and lightweight error correction for low-precision posit neural networks in safety-critical systems
Sadiq, S. and Lester, M.
It is advisable to refer to the publisher's version if you intend to cite from this work. See Guidance on citing. Abstract/SummaryLow-precision neural networks are increasingly adopted in safety-critical systems where energy efficiency and computational speed are essential. Posit number systems, particularly posit8 representations, offer favorable trade-offs between dynamic range and precision, but their resilience to hardware faults remains underexplored. This paper investigates the fault robustness of posit8 neural networks in a safety-critical advisory setting using the Posit Flight Advisory Network (PFAN) model. We develop a fast bit-flip fault injection framework targeting posit-encoded weights and evaluate advisory failure rates under single-bit, adjacent, and burst faults across both low and high-density fault injection scenarios. To mitigate fault-induced errors, we integrate two protection mechanisms: Hamming (13,8) SEC-DED codes for weight-level correction and output-level Triple Modular Redundancy (TMR). Combined ECC and TMR protection eliminates all observed advisory errors under single-replica fault conditions, and reduces the residual failure rate under double-replica corruption to 0.08%. Zero-failure configurations yield a 95% confidence upper bound of 2.16 failures per hour assuming a 10 Hz inference rate. These results demonstrate that lightweight protection strategies can enable dependable deployment of posit-based neural networks in safety-critical AI applications.
Deposit Details University Staff: Request a correction | Centaur Editors: Update this record |