Concurrent Self-Testing and Uncertainty Estimation of Neural Networks Using Uncertainty Fingerprint

Neural networks (NNs) are increasingly used in always-on safety-critical applications deployed on hardware accelerators employing various memory technologies. Reliable, continuous operation of NN is essential for safety-critical applications. During online operation, NNs are susceptible to (single and multiple) permanent and soft errors due to factors such as radiation, aging, and thermal effects. Explicit testing methods for hardware accelerators cannot detect transient faults during inference, are unsuitable for always-on applications, and require extensive test vector generation and storage. Therefore, in this paper, we propose the <italic>uncertainty fingerprint</italic> approach that represents the online fault status of NN. Furthermore, we propose a dual-head NN topology specifically designed to produce uncertainty fingerprints and the primary prediction of the NN in <italic>a single shot</italic>. During the online operation, by matching the uncertainty fingerprint, we can concurrently self-test NNs with up to 100% coverage in the backbone and in the full model (with a modified approach) with a low false positive rate while maintaining the performance of the primary task similar to the baseline. Compared to existing works, memory overhead is reduced by up to 243.7 MB, multiply and accumulate (MAC) operations are reduced by up to <inline-formula><tex-math notation="LaTeX">$10000\times$</tex-math><alternatives><mml:math><mml:mrow><mml:mn>10000</mml:mn><mml:mo>×</mml:mo></mml:mrow></mml:math><inline-graphic xlink:href="ahmed-ieq1-3640505.gif"/></alternatives></inline-formula>, and false-positive rates are reduced by up to 89%.