| dc.description.abstract | Despite current AI's human-like behavior, super efficiency, and unbelievable ability to handle clearly-defined complex tasks, it shows no sign of creativity, originality, or novelty outside its training set when facing tasks that are fundamentally different from but connected to its training set. Another aspect of this problem is that current AI is incapable of abstraction and reasoning in a generalizable, flexible, and systematic way. These abilities are closely related to core human cognitive factors, such as fluid intelligence and eductive ability. Inspired by how these abilities are examined in cognitive science, I put my dissertation in the context of visual abstract reasoning (VAR) tasks, a type of tasks widely used in human intelligence tests to measure core cognitive factors, and develop cognitive models, problem-solving systems, and learning models for solving VAR tasks. Two features make my computational models different from others. First, my computational models are inspired by mental imagery of human cognition, which is an type of imagistic representation that can be manipulated mentally. Some psychologists argue that mental imagery is a basic representation upon which other higher cognitive abilities are built, and thus important for creative visual thinking and reasoning in unfamiliar situations. Second, the information-processing and decision-making methods are deeply inspired by human’s analogy-making ability, which is argued to be the core of human cognition. From the classical analogy-making theories, I make one more step forward by extending them to a new analogy-making theory---consistency-based analogy making---and transform analogy-making processes into formal optimization processes which could be implemented with artificial neural nets. Besides these cognitively-inspired computational models, a comprehensive literature review for VAR tasks and computational models since 1930s and an outlook for applying generative AI on VAR tasks are also included in the dissertation. | |
