Feedforward and feedback sources of choice probability in neural population responses

Abstract

How the processing of signals carried by sensory neurons supports perceptual decisions is a long-standing question in neuroscience. The ability to record neuronal activity in awake animals while they perform psychophysical tasks near threshold has been a key advance in studying these questions. Trial-to-trial correlations between the activity of sensory neurons and the decisions reported by animals ('choice probabilities'), even when measured across repeated presentations of an identical stimulus provide insights into this problem. But understanding the sources of such co-variability between sensory neurons and behavior has proven more difficult than it initially appeared. Below, we discuss our current understanding of what gives rise to these correlations.

Figure 1

A simple scheme that produces the structured correlations required for CP in pooling models. The response of each neuron on any trial is the sum of two terms – a noise term that is independent for each neuron, and a common input that is the same for each neuron within a pool. The common inputs produce uniform positive noise correlations between pairs of neurons within a pool, but no correlation between pools. As N becomes large the pooled signals become dominated by the common input terms, since the independent noise terms tend to cancel. Thus the choices of the model (whether pooled up > pooled down) are largely determined by the common input terms. These therefore also determine CP in the model neurons. This property of linear pooling models remains true regardless of what gives rise to the common input terms. Whether these reflect noise in afferent neurons, or feedback from higher areas makes no difference.

Figure 2

Positive noise correlations need not limit the information available in a single pool, provided there is heterogenous tuning. A Consider a population of upward preferring neurons composed of two types: one with stronger rate modulation (type A) and the other weaker (type B). A common input to the pool (both type A and type B) generates uniform positive correlations between all pairs (see Fig. 1). The right panel illustrates across different trials of two stimulus strengths (S1 and S2), showing a high correlation between the trial-by-trial responses of the two neurons (blue and green circles, respectively). However, taking the difference of the responses between neuron A and B (bottom panel, black circles) removes this correlated noise, without losing the information about the stimulus. For illustration, a high correlation between the two neurons was used. Note that if the trial responses here represented the sum over a large population of type A neurons and type B neurons, the correlation would indeed approach 1 as pool size increases. This is because mean is dominated by the common input. The tuning heterogeneity means this subtraction produces a signal whose mean still depends on stimulus strength, so that here the ratio signal/noise increases with the number of neurons, with no upper bound. This is illustrated in panel B, where Fisher information is plotted for increasing homogeneous neuronal populations (left) and heterogeneous neuronal populations (right). The two top panels schematically depict the tuning curves in the respective populations (modified, with permission, after Ecker et al. 2011). C The situation when even an optimal linear decoder cannot remove the effect of correlated noise is illustrated with only two neurons. Left panel: the tuning curves of two neurons to the task relevant stimulus are shown. Right panel: The blue line depicts the response trajectory that corresponds to changes of the stimulus along the task-relevant dimension. If the correlated noise affects the population response along this trajectory, it is information limiting, since it cannot be differentiated from changes in the stimulus.