The most informative neural code accounts for population heterogeneity

Sensory perception depends on the accurate representation of the external world by neuronal populations. Each neuron’s contribution to a population code depends on its tuning preferences, but neuronal populations are heterogeneous: their spiking rate, spiking variability and tuning selectivity are all different. To understand how this heterogeneity impacts a neuron’s contribution to the population code, we recorded responses to moving stimuli from motion-sensitive neurons in anaesthetized marmosets (Callithrix jacchus) and trained linear decoders to discriminate direction using these responses. The relationship between a neuron’s preferred direction and the discrimination boundary was the strongest predictor of its decoding weight, and highly direction selective neurons are the most useful in the population code. Millisecond-timescale changes in neuronal responses mean that optimal weights change rapidly; however, perceptual readouts do not have this fine-grained temporal flexibility, and must perform sub-optimal decoding at each point in time. Author Summary It is widely understood that perception depends upon interpreting, or decoding, the pattern of activity across a neuronal population. To do this, firing rates from hundreds of sensory neurons must be weighted and combined to accurately represent stimulus properties. Despite the reality of a diverse, heterogeneous population of neurons, all we know is that their weighting is a function of the relationship between their tuning preference, and the decoding decision or perceptual task under consideration. In this manuscript we ask: which neurons in a heterogeneous population are the most informative, can their usefulness can be predicted by their response properties, and to what degree can we expect them to be optimized? We addressed this question using a machine learning approach to learn how to weight each neuron’s response in order to perform a range of discrimination tasks, and fit a series of decoding models to show that models based on preferred direction alone are significantly improved if neurons that are more selective are weighted more strongly. These results show that we can perform better decoding without machine learning if we address the false assumption that all members of the neural population are equally informative. Competing interests The authors have declared that no competing interests exist. Funding This work was funded by Australia’s National Health and Medical Research Council, the Human Frontier Science Program, and the Australian Research Council. Author Contributions EZ and NSCP designed the study, EZ, HHY, MGPR, and NSCP ran the experiments, EZ performed the analysis and modelling, EZ and NSCP wrote the paper.