Compressive learning of knowledge network

Poster Session B - Sunday, April 14, 2024, 8:00 – 10:00 am EDT, Sheraton Hall ABC

Muzhi Wang^1,2,3, Xiangjuan Ren^1,2,6,7, Tingting Qin⁴, Aming Li^4,5, Huan Luo^1,2,3; ¹School of Psychological and Cognitive Sciences, Peking University, ²PKU-IDG/McGovern Institute for Brain Research, Peking University, ³Beijing Key Laboratory of Behavior and Mental Health, Peking University, ⁴Center for Systems and Control, College of Engineering, Peking University, ⁵Center for Multi-Agent Research, Institute for Artificial Intelligence, Peking University, ⁶Max Planck Research Group NeuroCode, Max Planck Institute for Human Development, ⁷Institute of Psychology, Universität Hamburg

Knowledge is acquired by learning how isolated data in a network relate to each other, yet typical learning relies on inefficient random-walk explorations. We propose that higher-order associations in networks serve a key role in conveying relational knowledge and could be leveraged to facilitate network learning. We examined the hypothesis in two large-scale behavioral experiments and one magnetoencephalogy (MEG) study. Human subjects learned the transitional probabilities among 16 images by trial and error, with the transitional links comprising lattice, random, small-world, or scale-free networks. First, the scale-free network, endowed with strong inhomogeneous higher-order properties, displayed the highest learnability. Second, we developed a novel pre-learning HubToLeaf path that schematizes inhomogeneous higher-order network properties to facilitate subsequent network learning. Third, the HubToLeaf pre-learning path elicits stronger neural representations of the learned network in human brains, encoded in the anterior cingulate cortex (ACC). Finally, we built a computational model incorporating hypergraph theories to characterize higher-order structures and their impacts on network learning. Taken together, network learning benefits from higher-order network structures, emphasizing which would establish a 'compressive' scaffold for knowledge networks to develop.

Topic Area: EXECUTIVE PROCESSES: Working memory