Abstract
Performance generalization from low-to-high order tasks is crucial for abstract rules learning and decision-making. However, the computational understanding and further the neurocognitive mechanisms behind it are still elusive. Using three tasks involved in shape perception, emotional perception, and economic gambling combined with a unifying drift-diffusion model (DDM), we sought to explore the cognitive mechanisms of decision generalization from a neurocomputational perspective. Three parameters (drift rate, threshold, and non-decision time) were used as free parameters to infer the hidden decision/cognitive components at both group and individual levels after constraining the initial bias and sensory noise. Further correlation analyses based on individuals’ information sampling (or evidence accumulation) rate across three tasks were performed to identify the possibility of performance generalization. Behaviorally, we found a significant positive correlation (R=0.607, P<0.001) between emotion and shape task based on drift rate, suggesting generalization of the decision process in the same (perceptual) domain regardless of stimuli category. In addition, only a weak correlation (R=0.229, P=0.049) between shape and gambling task, and no correlation (R=0.159, P=0.15) between emotion and gambling task were observed, suggesting the independence of decision processes across perceptual and economic domains. Lastly, the parameters from the shape matching task can predict behavioral performance (i.e., reaction time) in the face-matching task at an individual level, further indicating performance generalization within the perceptual domain. Based on these, a representation similarity analysis (RSA) was used to capture the brain regions responsible for performance generalization while dynamic causal modeling (DCM) was introduced to identify the directional pathway and further address the hierarchical organization of the brain for decision generalization. Our preliminary results showed that the prefrontal-striatal network is mainly responsible for the higher-order cognitive performance generalization. Our findings may pave a new way for understanding the intrinsic functioning of the human brain and artificial general intelligence.