Value and choice as separable, stable representations in orbitofrontal cortex

Abstract

Value-based decision-making operates on multiple variables-including offer value, choice, expected outcome, and recent history-each functioning at different times in the decision process. Orbitofrontal cortex (OFC) has long been implicated in value-based decision-making, but it is unclear how downstream circuits might read out complex OFC responses into separate representations of the relevant variables to support different cognitive functions at specific times. We recorded from single neurons in OFC while macaque monkeys made cost-benefit decisions to juice offers. Using a novel analysis-optimal targeted dimensionality reduction-we discovered orthogonal, static dimensions (i.e. linear combinations of neurons) that selectively represented the value, choice, and expected reward of the present and, separately, previous offers. The neural composition of most representations was stable over discrete time periods that aligned to concurrent cognitive demands. We applied a new set of statistical methods to determine that the sensitivity, specificity and stability of the representations were greater than expected from the low-level features-dimensionality and temporal smoothness-of the responses alone. The separability and stability of OFC representations suggest a mechanism by which downstream circuits can read out specific task-relevant variables at appropriate times. Kimmel et al., 12/31/19 OFC Value and Choice Representations p. 2 of 103where 0 ≤ ≤ 1 specified the maximal accept rate, or saturation point of the psychometric curve. This parameter has previously been used to model the lapse rate, or intrinsic failure rate, of behavior 26 . The exponent took the form:where 3 was a constant and + determined the influence of the offered benefit [0, 1, 2, 4, 8] on trial t raised to , an exponent typical in economic models to implement a non-linear utility function 27 that was either fixed ( = 1) or allowed to vary as a free parameter. The benefit predictor was scaled to the range [0, 1]. Anatomy Supplementary Figure 4. MRI localization.Anatomical MRI sequences were acquired with recording grid in place prior to physiological recording. (a,b) T2-or T1-weighted coronal slice is shown at the median anterior-posterior extent of recorded sites for monkey N (a) or K (b), respectively. Animal's right side is shown on the left side of the image, per radiological convention. The inset shows an axial slice through the recording grid (white horizontal arrow), with a representative grid hole circled corresponding to grid position 2A or 0G (43.1 or 36.6 mm anterior to the interaural line, measured via intra-surgical stereotaxic coordinates; 6.7 or 6.3 mm left of midline, measured from the post-surgical MRI) for monkey N or K, respectively (see Supplementary Figure 7). Recording grid was filled with salinized agarose solution to facilitate contrast on MRI. By registering the axial slice through the recording grid and the coronal slice containing the selected grid position, we traced the virtual electrode trajectory from the bottom of...

Sequential and efficient neural-population coding of complex task information

“…S8B, red vs. black traces), albeit not for all variables because some variables such as position are highly non linearly coded when viewed at a global as opposed to local level. Our results are compatible with previous reports of long-timescale structure in neural state transitions in mouse posterior parietal cortex (Morcos and Harvey 2016b), long-timescale order in single-trial neural state trajectories in mouse premotor cortex (Wei et al 2019), and timescales of stability of neural representations in monkey orbitofrontal cortex (Kimmel et al 2020) (discussed further below). Strikingly, we observed that the efficient-coding-like function of the neural encoding geometry changed in time to approximately cancel out task variable correlations that were specific to each timepoint in the trial (Fig.…”

“…For neuron i in a multiplicative neural sequence, the proportion of activity variance explained by each task variable in is determined by the task-variable response function , which has no explicit time dependence. This is consistent with how the selectivity of neural representations can appear stable on longer timescales than the individual neural time-variations g i ( t ) (Kimmel et al 2020) (assuming that the measure of selectivity used is insensitive to overall activity scales as set by g i ( t )). Population-level information about one task variable relative to another, e.g.…”

Behavioral read out of population value signals in primate orbitofrontal cortex

“…To test this hypothesis, we recorded from populations of of OFC neurons in monkeys making value-based choices, and correlated their decisions with value representations calculated at the population level. Consistent with prior studies using multi-neuronal methods in OFC, we identify population-level representations of economic value 11,21,23,27 ; however, unlike prior work, we also identify a robust relationship between these value representations and the decisions made in individual trials.…”

“…These choice probability effects -derived from value representations measured at the population-level -are also substantially higher than those reported for single OFC cells in prior studies [22][23][24] . To directly compare population-based choice probabilities to single cells in our own study, we focused on the cells with non-zero weights in Eqns.…”

“…11 However, while these studies suggest some link between value representation and decision behavior, one key finding has been conspicuously absent: value-based choice probability. Even in large and rigorously analyzed samples of OFC cells, a significant trial-to-trial correlation between value coding and value-based choices has yet to be found [22][23][24] . This is especially puzzling; if OFC value signals mediate value-based choices, there must be some measurable relationship between them.…”