Final Report Summary - PREDISPIKE (Spike-based predictive coding: Closing the loop between neural dynamics and computation)
The vertebrate brain is presumably the most sophisticated information processing system in existence. However, while experimental recorded techniques improved exponentially (e.g. thousands of neurons recorded at the same time) our understanding of how brain activity controls cognition and behavior lags somewhat behind. Despite the huge heterogeneity and variability in single neural responses, these large recordings increasingly challenge the previously dominant “rate coding” assumption (i.e. that only the mean responses of largely redundant neural populations carry meaning or can impact behavior) without offering a clear alternative.
In this project we took an opposite view and considered how precisely tuned circuits of spiking neurons could learn and compute while spending minimal metabolic energy. The resulting circuits achieve single spike precision, while accounting for fundamental properties of neural activity (particularly its variability) and synaptic plasticity. This suggests that the brain learns and compute by consistently monitoring and correcting its own errors, by the way of maintaining a tight balance between excitatory and inhibitory currents, a principle that can be applied at multiple levels from single neurons to the entire brain. We used our approach to develop new methods for functionally interpreting large neural datasets. We also explore how disruptions in excitatory/inhibitory balance could lead to aberrant representations in the human brain and applied our framework to schizophrenia and bi-stable perception.
In this project we took an opposite view and considered how precisely tuned circuits of spiking neurons could learn and compute while spending minimal metabolic energy. The resulting circuits achieve single spike precision, while accounting for fundamental properties of neural activity (particularly its variability) and synaptic plasticity. This suggests that the brain learns and compute by consistently monitoring and correcting its own errors, by the way of maintaining a tight balance between excitatory and inhibitory currents, a principle that can be applied at multiple levels from single neurons to the entire brain. We used our approach to develop new methods for functionally interpreting large neural datasets. We also explore how disruptions in excitatory/inhibitory balance could lead to aberrant representations in the human brain and applied our framework to schizophrenia and bi-stable perception.