Closing The Loop: A Dynamic Neural Network Model Integrating Decision making and Metacognition

Abstract

Decision making under uncertainty entails selecting optimal actions from noisy evidence. Currently influential models, such as drift-diffusion and attractor frameworks, posit decisions as bottom-up stochastic evidence accumulation from sensory inputs but ignore critical interactions between decision, commitment, and metacognition. While these models explain basic choice behaviors and accompanying confidence, they fail to reconcile many empirical findings, including motor-area encoding of decision variables and time-dependent urgency signals. We present a closed-loop neural network model unifying three modules: decision module accumulating evidence, a motor module implementing action thresholds, and a metacognition module regulating deliberation through dual feedback pathways to suppress noise-driven errors and accelerate decision commitment under time constraints, respectively. This architecture can account for crucial characteristics of decision-making and metacognition that were empirically observed. By integrating decision, motor, and metacognitive dynamics, our model provides a biologically grounded framework for optimal decision making, offering testable predictions for neural and behavioral studies.