Reducing Influence of Noise in Regression Problems on Memristor Computing-in-Memory Chips

Memristor computing-in-memory (CIM) technology has shown great promise for creating energy-efficient computing chips. However, noise issues on the CIM chip remain problematic due to non-ideal factors such as process variations in device manufacturing and environmental interferences, especially in regression problems. This work delves into the effects of noise on various regression models, encompassing linear regression and vector regression, as well as on the confidence value prediction in object detection models. We introduce novel calibration algorithms both average error and dynamic error as well as confidence threshold calibration techniques, specifically designed to mitigate noise induced inaccuracies. Comprehensive experimentation across diverse datasets validates the effectiveness of our proposed algorithms, highlighting their potential for enhancing CIM chip performance for regression tasks in the presence of noise.