Основні поняття
Neural dynamics minimize prediction errors by integrating confidence in cortical circuits.
Статистика
"Empirical studies on humans and other animals show that prior knowledge and data from multiple modalities are weighted by their relative uncertainty during perceptual integration [4], decision-making [5, 6] and sensorimotor control [7, 8]."
"Potential implementations in the cerebral cortex have been discussed, notably in cortico-pulvinar loops [25] or more generally through neuromodulation [26, 27]."
"In this work, we suppose that cortical areas must not only predict the activity in other areas and sensory streams but also jointly estimate the confidence of their predictions, where we define confidence as the (estimated) inverse expected uncertainty of the prediction."
"At the equilibrium of neuronal dynamics, weights of synapses carrying predictions can be learned following the gradient."
"Weights Aℓ of synapses carrying confidence can also be learned following the gradient."
Цитати
"In a forest, I know there are trees."
"Across trees, structure (trunk, branches, leaves, etc.) is usually more consistent than color."
"Having formulated an energy for cortical function, we formally derive gradient-based neuronal dynamics and synaptic learning rules minimizing this energy."