Weighting neurons by selectivity produces near-optimal population codes

Area of Science:

• Neuroscience
• Computational Neuroscience
• Sensory Coding

Background:

• Perception relies on reading out neural representations of sensory stimuli.
• Decoding neural population activity often involves weighted sums of neuronal spike counts.
• Variability in optimized weights for in vivo recordings complicates understanding their practical impact on decoding.

Purpose of the Study:

• To experimentally investigate how neuronal population activity codes sensory information.
• To determine which aspects of a neuron's tuning predict its weighting in decoding models.
• To assess the impact of different weighting schemes on decoding performance.

Main Methods:

• Recorded neuronal activity from the middle temporal area (MT) of anesthetized marmosets.
• Presented visual stimuli consisting of dots moving coherently in 12 different directions.
• Employed optimized decoding models (generalized linear model, Fisher's linear discriminant) and models with a priori weights based on neuronal tuning.

Main Results:

• High peak response and direction selectivity were strong predictors of higher weights in optimized decoding models.
• Learned weights differed significantly from a priori rules but resulted in only marginally better decoding performance.
• Models using a priori weights based on preferred direction and selectivity achieved near-maximal decoding performance.

Conclusions:

• Neuronal direction selectivity is a critical factor determining a neuron's contribution to sensory coding.
• Weighting neurons based on their selectivity causally improves population-level direction representation.
• Decoding performance is robustly predicted by neuronal tuning properties, with optimized weights offering only marginal benefits over selectivity-based weighting.