BATTCAVE Seminar Series

Sept 16, Thursday, 11:30-12:30, Fenglian Pan (UNC Charlotte, ISE)

Title: Computationally Efficient Learning of Artificial Intelligence System Reliability Considering Error Propagation

Abstract: Artificial intelligence (AI) systems are increasingly prominent in emerging smart cities. Even with their demonstrated promising potential, reliability assessment for AI systems is in high demand before large-scale deployment. Unlike traditional systems, AI systems typically fuse heterogeneous data streams and operate through a sequence of interdependent functional stages. Errors that occurred in upstream stage(s) may propagate and cause additional errors in downstream stage(s).  This error propagation will be accumulated and ultimately affect the reliability of the overall AI system. Understanding and quantifying the impact of such error propagation is critical, yet remains challenging due to three main factors: i) data availability: real-world AI systems reliability data are often scarce and constrained by privacy concerns; ii) modeling complexity: error events across sequential stages are interdependent, violating the independence assumptions of most existing reliability models; and iii) scalability limitations: AI systems process large volumes of high-speed data, resulting in frequent and complex recurrent error events that are difficult to track and analyze. To address these challenges, this paper proposes a systematic and justifiable error injection framework that incorporates real-world information into a physics-based simulation to efficiently generate reliable data. With the simulated data, a statistical model is formulated to explicitly consider the error propagation for AI systems reliability modeling. To estimate model parameters from extensive error event data, a computationally efficient and theoretically guaranteed composite likelihood expectation-maximization algorithm is proposed. Our approaches are applied to assess the reliability of AI systems in autonomous vehicles. Both numerical experiments and physics-based simulation case studies demonstrate the prediction accuracy and computational efficiency of the proposed methods.

Oct. 17, Friday, 11:30-12:30, Nan Zhao (UNC Charlotte, MEES)

Nov. 14, Friday, Christofer Bejger (UNC Charlotte, CHEM)

Dec. 12, Friday, 11:30-12:30, Youxing Chen (UNC Charlotte, MEES)

Jan. 16, Friday, 11:30-12:30, Farah Deeba (UNC Charlotte, ECE)

Feb. 13, Friday, 11:30-12:30 Sheldon Xie (UNC Charlotte, ETCM)

Mar. 13, Friday, TBD

Apr.  10, Friday, TBD

May. 8. Friday, TBD