2026-05-20T04:56:08Zhttps://keep.lib.asu.edu/oai/request

oai:keep.lib.asu.edu:node-2011242025-05-05T15:53:02Zoai_pmh:alloai_pmh:repo_items

201124 https://hdl.handle.net/2286/R.2.N.201124 http://rightsstatements.org/vocab/InC/1.0/ All Rights Reserved 2025 184 pages Doctoral Dissertation Academic theses en Hong, Jinyung Pavlic, Theodore P. Smith, Brian Baral, Chitta Yang, Yezhou Lee, Heewook Arizona State University Partial requirement for: Ph.D., Arizona State University, 2025 Field of study: Computer Science Despite recent advancements in Artificial Intelligence (AI), substantial progress is still needed for AI systems to match even simple organisms, such as fruit flies, in cognitive and sensorimotor abilities. Although state-of-the-art vision-language models and foundational systems exhibit novel emergent behaviors, they remain inferior to basic animal species in perception and reasoning. For example, honey bees can quickly learn general features of complex floral odors with very little data, while AI systems often require vast amounts of training to generalize effectively. My research explores biologically inspired strategies to enhance AI performance, demonstrating how cognitive frameworks can improve explainability and human-understandable reasoning. This thesis develops three novel approaches that integrate insights from neuroscience and cognitive science to advance AI learning, adaptability, and interpretability. First, I apply random projection, a key feature of insect-brain architecture, to enable efficient associative learning in deep-learning frameworks by reducing dimensionality while preserving important structural relationships and creating opportunities for modularity and reuse in size-constrained architectures. Second, inspired by neuromodulatory mechanisms that facilitate continual learning in insect brains, I designed a neural framework that connects task representations through geometric-sensitive hashing, allowing AI models to retain and transfer knowledge across tasks more effectively. Finally, drawing from Global Workspace Theory, a cognitive science framework modeling conscious information processing, I propose an AI architecture that enhances interpretability by structuring decision-making in a human-understandable manner, improving transparency and trust in AI systems. Surprisingly, these biologically inspired modifications not only improved performance but also led to more modular and explainable AI systems. This suggests that biological systems may inherently promote transparency, even though explainability is not their primary evolutionary function. Moreover, the enhanced interpretability exhibited by deep networks structured to mimic models of conscious thought suggests that consciousness itself might emerge as a byproduct of adaptive brain structures with deeper organizational benefits. These findings illustrate how insights from biology and cognitive science can reveal structural advantages overlooked in traditional AI design, which relies on arbitrary optimization. This work suggests that AI designed with biologically plausible mechanisms may naturally align with human-intuitive reasoning, offering new pathways toward more adaptable and trustworthy intelligent systems. Computer Science Putting the Mind and the Body into AI: Using Random Projections, Neuromodulation, and Cognitive Modules to Improve the Design of Modern AI Systems