Do Deep Learning Models Mimic Human Personality Traits? – An Empirical Study

Abstract

<p>Two key aspects of artificial intelligence are its ability to make decisions and attempt to mimic humans. Decision-making in humans is, however not straightforward and depends significantly on the person's mental state, personal biases, and personality. In this study, we attempt to empirically understand if deep learning image classifiers also exhibit such inherent biases or if they act neutrally in any given situation. To this end, we perform three experiments – left-brain right-brain test, psychological images test, and Rorschach's inkblot test on eight different stat-of-the-art deep learning classifiers. A detailed analysis of the SoftMax probability scores is done rather than an analysis on measures like accuracy and F1. The experimental results suggested that most models work similar to a left-brained person, do not always predict the same class when given images consisting of multiple object classes, and usually detect larger objects rather than smaller ones. We believe that understanding these inherent biases would help future researchers take necessary actions while building image classification models.</p>