Predictive coding of dynamical variables in balanced spiking networks.

Area of Science:

• Neuroscience
• Computational Neuroscience
• Neural Networks

Background:

• Cortical neuroscience faces two long-standing puzzles: high neural response variability and the precise balance of excitation and inhibition.
• Understanding these properties is key to deciphering neural computation.

Purpose of the Study:

• To demonstrate that neural variability and balanced excitation-inhibition are necessary consequences of efficient information representation in spiking neural networks.
• To propose a novel framework for understanding neural coding and computation in the brain.

Main Methods:

• Developed spiking neural networks (using leaky integrate-and-fire neurons) that represent dynamical variables efficiently.
• Assumed linear readout of information from spike trains and that spikes are fired only to improve representation.
• Derived network dynamics equivalent to prediction error signals.

Main Results:

• The derived networks implement arbitrary linear dynamical systems.
• Neuron membrane voltage corresponds to prediction error about a population-level signal.
• Constructed a robust integrator network of spiking neurons.
• Demonstrated that neural variability in these networks is not noise, leading to orders of magnitude greater reliability than traditional population coding models.

Conclusions:

• Neural variability and balanced excitation-inhibition are emergent properties of efficient spiking neural network coding.
• Spiking activity is fundamental to brain computation, challenging previous assumptions.
• The reliability of cortical representations may be significantly higher than currently estimated.