METHOD FOR DETECTING AND MITIGATING BIAS AND WEAKNESS IN ARTIFICIAL INTELLIGENCE TRAINING DATA AND MODELS

摘要

An exemplary embodiment may present methods for detecting bias both globally and locally by harnessing the white-box nature of the eXplainable artificial intelligence, eXplainable Neural Nets, Interpretable Neural Nets, eXplainable Transducer Transformers, eXplainable Spiking Nets, eXplainable Memory Net and eXplainable Reinforcement Learning models. Methods for detecting bias, strength, and weakness of data sets and the resulting models may be described. A first exemplary method presents a global bias detection which utilizes the coefficients of the explainable model to identify, minimize, and/or correct any potential bias within a desired error tolerance. A second exemplary method makes use of local feature importance extracted from the rule-based model coefficients to identify any potential bias locally. A third exemplary method aggregates the feature importance over the results/explanations of multiple samples. A fourth exemplary method presents a method for detecting bias in multi-dimensional data such as images. Further, a backmap reverse indexing mechanism may be implemented. A number of mitigation methods are also presented to eliminate bias from the affected models.