Reliability Assessment of Tiny Machine Learning Algorithms in the Presence of Control Flow Errors

Abstract

With the advances in hardware technologies, embedded and Edge devices are now able to offer sufficient memory and computational power to accommodate light-weight machine-learning (ML) classifiers. However, due to the intensive code optimization and summarization in the design phase, the reliability of light-weight ML applications is at risk. In this paper, we study the reliability of three prototypical light-weight ML applications against the well-known control flow (CF) errors. We have injected a total of 66,156 CF errors into Bonsai, ProtoNN, and TensorFlow Lite ML applications running on the Arduino board. Based on the results obtained from the error-injections, we found that CF errors could affect either the functionality or the classification accuracy of the ML-based inference as the embedded application. We conclude that making a single decision only after a long sequence of computations/branches may be more error-prone. This issue can be addressed by combining the inferences of intermediate nodes in the chain to obtain the final classification decision.