Observations and theories: cognitive mechanisms of discovery

Abstract

This dissertation research lies at the intersection of cognitive science, psychology, machine learning, and philosophy, aiming to enhance our understanding of how theories and data are—and should be—integrated to arrive at understandings of the world. In the first part, I empirically examine how our everyday concepts and theories shape our perception of the world. I further investigate the impact of scientific conceptual systems, such as the taxonomy of brain regions, on scientific experimentation. I conclude this part by conducting a computational study examining the effectiveness of different theory-motivated experimentation strategies at guiding agents towards useful theories of the world. In the second part, I explore the construction of theories based on observations, emphasizing the human and scientific bias towards relatively simple representations. I critique this preference for simpler accounts and introduce the concept of learning with excess capacity, or a complexity bias, as an alternative approach that allows learning systems to develop useful representations of the world in many contexts. Then, I build on this insight to re-examine the principle of model parsimony in science, identifying contexts where the preference for simpler scientific models could either facilitate or impede scientific progress. The third part broadens the scope by analyzing social aspects that influence the emergence and evolution of useful concepts in a community of interacting agents. Here, I identify the conditions that promote the development of stable communicative conventions, including the role of social network structure, supervision, and the strategy of starting small and gradually expanding one’s vocabulary. Finally, in part four I reflect on the mutual interactions between concepts and observations in science by drawing insights from empirical research on human concept learning and conceptual influences on human perception. I conclude by suggesting that progress in science can be facilitated by both empirical (with psychological experiments) and formal (with mathematical and computational models) examinations of the cognitive processes of observing and representing the world.