Human Perceptual Biases and their Analogous Prevalence in Artificial Intelligence Systems: Bias Mitigation for Dependable AI

Artificial Intelligence (AI) is entering and influencing all facets of our life. Unintended bias in these AI systems can have serious consequences and are of great concern. We have made tremendous progress in the field of AI by studying the various nuances of human intelligence spanning the metaphysical to the physiological perspectives. Deep Learning was inspired by the structure of neurons in the human brain. Similarly, a study of the shortcomings and fallibility of human intelligence will help us design dependable AI systems that are robust, reliable, and safe. With this in mind, this study uses findings from Psychophysics for some of the most common human perceptual biases in magnitude estimation (viz. Regression effect, Range effect, Scalar Variability, and Sequential effect) to check for their analogous prevalence in four Machine Learning (ML) models based on Support Vector Machines, LSTM, CNN-GRU, and Transformers, for stock market price prediction task. We found that the ML models, which are assumed to be objective and immune to emotions, can be as biased as humans. The paper discusses the implications of these biases in general and shows methods for their mitigation using data augmentation and synthetic data generated using Wasserstein GAN, which also suggests that a Bayesian framework may explain the presence of these biases.