AI today performs several functions in everyday life from chatbots that respond to tax questions to dish out medical diagnoses and algorithms that maneuvre autonomous vehicles. According to researchers at the University of California, it requires a hybrid human-machine technique to create a smarter, more accurate systems.
A study published in the Proceedings of the National Academy of Sciences describes a new mathematical model that can improve performance of AI by combining human and algorithmic projections and confidence ratings.
In fact, humans and machine algorithms exhibit complementary strengths and weaknesses. Each uses different strategies and different sources of information to anticipate and to make decisions, stated the co-author of the study.
Meanwhile, it is shown using observed evidence and theoretical analysis, humans can improve the estimation of AI. This is true even when human accuracy is somewhat below than that of AI and vice versa. Importantly, this accuracy exceeds than the one experienced when predictions from two individuals or two AI algorithms are combined.
In order to validate this, researchers carried out an image classification experiment. This involved human participants and computer algorithms to work separately for the correct identification of distorted pictures of animals and everyday objects such as chairs and trucks. On observation, human participants rated their confidence in the accuracy of each image as low, high, or medium, whilst the machine classifier created a continuous score. The results had large differences in the confidence between AI algorithms and humans across images.
To understand the difference, for example, human participants were quite confident for a particular picture to contain a chair, while AI algorithm was not sure of this. To validate this further, for some other images, AI algorithm was able to confidently provide a label for the object under display, which human participants were not sure of.