What Really Matters is not just Knowing “What”, “Where” and “When” but also Knowing “How”

Abstract

Human intelligence has been an object of investigation since the beginning of the research on information science to provide artificial agents with human-like decision making skills. This research field has led to the development of algorithms that try to simulate human reasoning. Several theories have been defined to model decisions in the presence of uncertain, imprecise and vague information, based on both subjective and qualitative criteria, expressed linguistically. Today, we are at an epochal turning point in which there are no longer attempts to reproduce human reasoning by machines, but algorithms are designed as networks of interconnected simple computational units learning to take decisions from examples. This data-driven paradigm simulates children learning from observations, so that their behavior evolves by accumulation of experience. Nevertheless, are we sure that purely learning from data is an effective sufficient method, not affected by bias, and that it can lead to fair systems that we can trust? Are we satisfied with completing a task without knowing “how” it was performed? Are we sure that children don’t have, a priori, more complex and structured mechanisms regulating as well as directing their learning ability? Do we really want to throw away all the research that has been done so far, or can we retain it, so that knowledge of models and data-driven learning can play a synergistic role?