Biases in neural population codes with a few active neurons.

Area of Science:

• Neuroscience
• Computational Neuroscience
• Information Theory

Background:

• Neural information is processed via population codes, involving simultaneous activity of multiple neurons.
• Decoding accuracy is often assessed by trial-to-trial variations in stimulus estimation, especially with noisy neurons.

Purpose of the Study:

• To characterize the bias in population codes when few neurons are active.
• To investigate the influence of encoding models and noise on decoding biases.
• To introduce a method for estimating bias and variance in Bayesian decoders.

Main Methods:

• Analysis of encoding models, including rectified cosine tuning and von Mises functions.
• Evaluation of biases in maximum likelihood and Bayesian decoders.
• Development of a technique to estimate bias and variance for Bayesian least squares decoders.

Main Results:

• Biases emerge when only a few neurons are active, leading to systematic differences between true and estimated stimuli.
• The shape of the bias is dependent on the encoding model and can be attractive or repulsive.
• Decoding biases exhibit a non-trivial dependence on neural noise levels.

Conclusions:

• Understanding and quantifying bias is crucial for accurate interpretation of neural population codes, particularly in scenarios with limited neuronal activity.
• The introduced technique provides a valuable tool for analyzing decoder performance in low-activity neural populations.