Human Orbitofrontal Cortex Represents a Cognitive Map of State Space

Abstract

Although the orbitofrontal cortex (OFC) has been studied intensely for decades, its precise functions have remained elusive. We recently hypothesized that the OFC contains a "cognitive map" of task space in which the current state of the task is represented, and this representation is especially critical for behavior when states are unobservable from sensory input. To test this idea, we apply pattern-classification techniques to neuroimaging data from humans performing a decision-making task with 16 states. We show that unobservable task states can be decoded from activity in OFC, and decoding accuracy is related to task performance and the occurrence of individual behavioral errors. Moreover, similarity between the neural representations of consecutive states correlates with behavioral accuracy in corresponding state transitions. These results support the idea that OFC represents a cognitive map of task space and establish the feasibility of decoding state representations in humans using non-invasive neuroimaging.

Copyright © 2016 Elsevier Inc. All rights reserved.

Figures

Figure 1. Experimental task

(A): Example of trial sequence in the task. Participants began by judging the age (young vs. old) of the cued category (face or house). In the following trials, they continued to judge the age of the same category until an age change occurred. On the trial following an age change, participants had to switch to judging the other category, with the first trial after the change determining the new age of that category to which subsequent trials needed to be compared. These rules created an alternating mini-block structure of judging either the age of faces or houses. The first trial after a category switch was the trial in which the mini-block was entered (Enter trial), trials in which the category and age repeated are denoted Internal trials, and the trial in which the age changed was the end of the current mini-block (Exit trial). (B): Trial structure. Each trial started with a fixation cross and the display of the randomly-determined response mapping for the current trial. Following the stimulus display, participants had up to 2750 ms to make their response before the next trial started, while the stimulus duration was independent of the response time and lasted on average 3300 ms. Responses were followed by a box around the chosen option. Wrong responses led to a repetition of the same (Enter trials) or preceding trial (Internal and Exit trials), accompanied by a written reminder of the current category. (C): Mental operations involved in different trial types. The diagram illustrates how currently hidden information about the previous age as well as the current and previous category must be factored into the decision-making process. (D): Possible transitions between states during the task. Each circle denotes a particular state (see legend). Arrows indicate possible transitions and node colors indicate the trial types.

Figure 2. Encoding of hidden state information in OFC BOLD signals

(A): Average 16-way classification of state identity from fMRI patterns within the anatomically defined OFC (blue bar) and following a permutation test (black bar). (B): Contribution of information to 16-way classification separately for task-irrelevant information from 2 trials ago (leftmost bars, light blue), the three different hidden state components (solid blue) as well as the observable state component (striped bar). Only hidden and task-relevant components of the state contributed significantly to state identity decoding. (C): Dendrogram indicating the similarity structure of different states according to a hierarchical cluster analysis. Colors and acronyms as in Figure 1D. Note that all purple (Enter) states involve a category switch, whereas all other states do not. (D): Decoding of hidden state information separately for switch and non-switch trials. Dashed horizontal lines: chance baseline, error bars: S.E.M, *’s: p ≤ .05.

Figure 3. Anatomical specificity of hidden state representations

(A–D): searchlight maps for decoding accuracy (t-test against chance baseline) for each of the four state components (A) previous category, (B) previous age, (C) current category and (D) current age, thresholded at p = .025 (uncorrected) for illustration purposes. Colors represent t-values. Sagittal, coronal and axial slices are at x = 4, y = 44 and z = −14 (MNI) respectively. (E): Conjunction analysis of three hidden components, showing that all three hidden components could be decoded simultaneously only within medial OFC (conjunction threshold p < .01, i.e., every state component is significant with p ≤ .01, uncorrected, at shown location). Red voxels indicate significant searchlight centers; blue line indicates the outline of the corresponding searchlights. A complete table of results can be found in SI, Table S1.

Figure 4. OFC state classification correlates with task performance

(A): A significant relationship between average classification accuracy within OFC (mean across all three hidden state components) and participants’ error rate suggests that the encoded state information was relevant for task performance. Each dot represents one participant. Red: regression line. (B): Trialwise decoding before, during and after behavioral errors. Empty circles: 16-way decoding accuracy in the 5 trials leading up to, during and after an error. Filled circles: decoding accuracy during 7 consecutive trials with no behavioral error. Chance = 6.25%, error bars: S.E.M.

Figure 5. OFC state representations affect task performance

(A): Average correlations between neural state representations within OFC. Darker gray denotes higher correlation (i.e., more similar state representations). (B): Relationship between error rate on the 8 different transitions exiting one miniblock and entering another and correlation between the pairs of states corresponding to these transitions, across participants. For each participant, the 8 transitions were ordered according to strength of correlation (from low to high). Dots denote the average correlation between states in that ordinal position across participants (x axis) and average behavioral error rate on the corresponding transitions (y axis), with horizontal and vertical error bars denoting S.E.M of each. Higher correlations between neural states were associated with fewer behavioral errors, on average (p = .04). (C): Histogram of within-subject correlations between error rates and neural state similarity showing that correlations were, on average, significantly lower than 0 (p < .01).