Stimulus dependent relationships between behavioral choice and sensory neural responses

Abstract

Understanding the relationship between trial-to-trial variability in neural responses of sensory areas and behavioral choices is fundamental to elucidate the mechanisms of perceptual decision-making. In two-choice tasks, activity-choice covariations have traditionally been quantified with choice probabilities (CP). It has been so far commonly assumed that choice-related neural signals are separable from stimulus-driven responses, which has led to characterizing activity-choice covariations only with a single CP value estimated combining trials from all stimulus levels. In this work we provide theoretical and experimental evidence for the stimulus dependence of the relationship between neural responses and behavioral choices. We derived a general analytical CP expression for this dependency under the general assumption that a decision threshold converts an internal stimulus estimate into a binary choice. This expression predicts a stereotyped thresholdinduced CP modulation by the stimulus information content. We reanalyzed data from and found evidence of this modulation in the responses of macaque MT cells during a random dot discrimination task. Moreover, we developed new methods of analysis that allowed us to further identify a richer structure of cell-specific CP stimulus dependencies. Finally, we capitalised on this progress to develop new generalized linear models (GLMs) with stimulus-choice interaction terms, which show a higher predictive power and lead to a more precise assessment of how much each neuron is stimulus-or choice-driven, hence allowing a more accurate comparison across areas or cell types. Our work suggests that characterizing the patterns of stimulus dependence of choice-related signals is essential to properly determine how neurons in different areas contribute to linking sensory representations to perceptual decisions.

Stimulus-dependent relationships between behavioral choice and sensory neural responses

“…This shows that in fact the use of corrected z-scores to pool responses across stimulus levels to calculate a grand CP is in the linear approximation equivalent to calculating a weighted average of the CPs at each stimulus level, with a specific selection of the average weights, namely . An analogous derivation with the uncorrected z-scoring shows that in that case the pooling across stimulus levels is associated with an improper use of unormalized weights, which analytically confirms the arguments and simulations in Kang and Maunsell, 2012 indicating that a grand CP calculated with the standard z-scoring provides a biased estimation of an underlying stimulus independent CP (see a detailed derivation in section S2 of Chicharro et al, 2019 ). The weights differ from the ones inversely proportional to the standard error of the CP estimates ( Equation 18 ).…”

Large-scale dynamics of perceptual decision information across human cortex

“…Both assumptions are simplified: contrast response functions in MEG power may not be perfectly linear in all individuals (see ref. 45 for demonstration for motion coherence response functions), and stimulus-and choice-related cortical signals may interact multiplicatively 46 . Consequently, the absence of a choice-predictive component in residual V1 gammaband activity observed in our analysis may reflect technical limitations, such as a failure of our linear approach to isolate the endogenous fluctuations and/or the low signal-to-noise ratio of MEG gamma-band activity 33 .…”