Skip to main content

Main menu

  • HOME
  • CONTENT
    • Early Release
    • Featured
    • Current Issue
    • Issue Archive
    • Collections
    • Podcast
  • ALERTS
  • FOR AUTHORS
    • Information for Authors
    • Fees
    • Journal Clubs
    • eLetters
    • Submit
    • Special Collections
  • EDITORIAL BOARD
    • Editorial Board
    • ECR Advisory Board
    • Journal Staff
  • ABOUT
    • Overview
    • Advertise
    • For the Media
    • Rights and Permissions
    • Privacy Policy
    • Feedback
    • Accessibility
  • SUBSCRIBE

User menu

  • Log out
  • Log in
  • My Cart

Search

  • Advanced search
Journal of Neuroscience
  • Log out
  • Log in
  • My Cart
Journal of Neuroscience

Advanced Search

Submit a Manuscript
  • HOME
  • CONTENT
    • Early Release
    • Featured
    • Current Issue
    • Issue Archive
    • Collections
    • Podcast
  • ALERTS
  • FOR AUTHORS
    • Information for Authors
    • Fees
    • Journal Clubs
    • eLetters
    • Submit
    • Special Collections
  • EDITORIAL BOARD
    • Editorial Board
    • ECR Advisory Board
    • Journal Staff
  • ABOUT
    • Overview
    • Advertise
    • For the Media
    • Rights and Permissions
    • Privacy Policy
    • Feedback
    • Accessibility
  • SUBSCRIBE
PreviousNext
Research Articles, Behavioral/Cognitive

Neural Representations in the Prefrontal Cortex Are Task Dependent for Scene Attributes But Not for Scene Categories

Yaelan Jung and Dirk B. Walther
Journal of Neuroscience 25 August 2021, 41 (34) 7234-7245; https://doi.org/10.1523/JNEUROSCI.2816-20.2021
Yaelan Jung
Department of Psychology, University of Toronto, Toronto, Ontario M5S 3G3, Canada
  • Find this author on Google Scholar
  • Find this author on PubMed
  • Search for this author on this site
  • ORCID record for Yaelan Jung
Dirk B. Walther
Department of Psychology, University of Toronto, Toronto, Ontario M5S 3G3, Canada
  • Find this author on Google Scholar
  • Find this author on PubMed
  • Search for this author on this site
  • ORCID record for Dirk B. Walther
  • Article
  • Figures & Data
  • Info & Metrics
  • eLetters
  • PDF
Loading

Abstract

Natural scenes deliver rich sensory information about the world. Decades of research has shown that the scene-selective network in the visual cortex represents various aspects of scenes. However, less is known about how such complex scene information is processed beyond the visual cortex, such as in the prefrontal cortex. It is also unknown how task context impacts the process of scene perception, modulating which scene content is represented in the brain. In this study, we investigate these questions using scene images from four natural scene categories, which also depict two types of scene attributes, temperature (warm or cold), and sound level (noisy or quiet). A group of healthy human subjects from both sexes participated in the present study using fMRI. In the study, participants viewed scene images under two different task conditions: temperature judgment and sound-level judgment. We analyzed how these scene attributes and categories are represented across the brain under these task conditions. Our findings show that scene attributes (temperature and sound level) are only represented in the brain when they are task relevant. However, scene categories are represented in the brain, in both the parahippocampal place area and the prefrontal cortex, regardless of task context. These findings suggest that the prefrontal cortex selectively represents scene content according to task demands, but this task selectivity depends on the types of scene content: task modulates neural representations of scene attributes but not of scene categories.

SIGNIFICANCE STATEMENT Research has shown that visual scene information is processed in scene-selective regions in the occipital and temporal cortices. Here, we ask how scene content is processed and represented beyond the visual brain, in the prefrontal cortex (PFC). We show that both scene categories and scene attributes are represented in PFC, with interesting differences in task dependency: scene attributes are only represented in PFC when they are task relevant, but scene categories are represented in PFC regardless of the task context. Together, our work shows that scene information is processed beyond the visual cortex, and scene representation in PFC reflects how adaptively our minds extract relevant information from a scene.

  • fMRI
  • MVPA
  • PFC
  • PPA
  • scene categorization
  • task dependency

Introduction

Real-world scenes provide a wealth of complex sensory experiences. Research has demonstrated that various components of scenes are processed across several brain regions of the scene-selective network in the visual cortex, such as the parahippocampal place area (PPA; Epstein and Kanwisher, 1998), the retrosplenial cortex (Maguire, 2001), and the occipital place area (Dilks et al., 2013). However, recent studies have shown that visual scene processing takes place beyond the visual cortex and involves the associative cortex such as the parietal (Silson et al., 2016, 2019) and the prefrontal cortex (PFC; Jung et al., 2018). How are scene categories and other scene attributes represented beyond the scene-selective network? In the current study, we investigate whether and how scene content is processed beyond the visual cortex, especially in PFC.

We have previously demonstrated that scene images and sounds elicit neural representations of scene categories in PFC. Intriguingly, these representations are modality independent, suggesting that PFC represents scene content in an abstract way, not constrained in a single sensory modality (Jung et al., 2018). Still, we do not know yet what elements of scene content are processed and represented in PFC. That is, natural scenes also comprise other scene attributes than just scene categories, some of which, based on spatial elements, are often referred to as global scene properties, such as openness and navigability (Greene and Oliva, 2009). Scene attributes also include nonspatial scene properties, such as overall temperature, transience, or functional affordances (Greene et al., 2016). Research has shown that many scene attributes are processed in scene-selective visual brain regions (Kravitz et al., 2011; Park et al., 2011; Persichetti and Dilks, 2019). However, we do not know yet whether PFC represents scene attributes that are not related to scene categories, and whether the neural presentations of scene attributes in PFC are also modality independent.

Scene representations in PFC may also depend on task context. One of the fundamental functions of the PFC is to modulate the neural representation of sensory inputs according to task demands (Chadick and Gazzaley, 2011; Zanto et al., 2011; Harel et al., 2014). Thus, representations of scene content in PFC might be dependent on their task relevance. However, research has shown that scene category processing is not necessarily impacted by tasks; behavioral studies have shown that scene category information is accessible with limited attention (Li et al., 2002), and even when it is not relevant to the task context (Greene and Fei-Fei, 2014). Furthermore, scene categories have been shown to be represented in PFC in a passive-viewing paradigm (Walther et al., 2009; Jung et al., 2018), which does not explicitly require attention to scene categories. Therefore, we predict that task context is more likely to impact the processing of scene attributes than that of scene categories. We explore these possibilities by assessing the neural representations of scene attributes and scene categories in PFC under different task conditions.

The current study uses two scene attributes (temperature and sound level) and four natural scene categories (beaches, cities, forests, and train stations) to examine how scene attributes and scene category information are represented in PFC across different task contexts. To construct different task conditions for each scene property, we instructed participants to make a judgment on a scene attribute, which would encourage them to attend to a particular property in a scene.

Finally, the present study also explores whether representations of scene attributes in PFC are modality independent in the same way as scene categories (Jung et al., 2018). We do so by comparing the neural representation of scene attributes (both temperature and sound level) inferred from images to the neural representation of direct thermal and auditory percepts. Our findings demonstrate that both scene attributes and scene categories are represented in PFC in a modality-independent way. Neural representations of scene attributes generalize across sensory modalities, as indicated by similar neural activity patterns across percepts inferred from images and direct sensation.

Materials and Methods

Participants

Twenty-five healthy adults recruited from the University of Toronto participated in the fMRI study (10 males, 15 females; mean age = 22.17 years; SD = 1.99; all right handed). A power analysis based on our previous study using a similar decoding technique (Jung et al., 2018) indicates that we would require 14 participants for similar power. As the current study uses an event-related design instead of a block design, we expected that our effect would be more subtle, and thus we would need more participants. We performed a pilot study with five participants to examine the feasibility of the decoding methods (data not included here). Based on the pilot data, we decided to collect data from 20 participants, provided that they complete both sessions of the experiment.

All participants reported normal or corrected-to-normal vision and normal hearing, and no history of neurologic abnormalities. Five participants did not complete both sessions, and their data were excluded from the analysis. Participants provided informed consent before the experiment, and all aspects of the experiment were approved by the Research Ethics Board at the University of Toronto (protocol #33720).

Experiment design and stimuli

The study consisted of two sessions: the main experiment session and the localizer session. For almost all participants, the main experiment session was performed first, except for three participants who performed the localizer experiment first for logistical reasons (e.g., last-minute cancellation). The main experiment session consisted of three different types of tasks: fixation monitoring, temperature judgment, and sound-level judgment. Participants always performed two runs of the fixation task first, and then the other two tasks in alternation (either T-S-T-S or S-T-S-T). The localizer session consisted of three different experiments: the sound experiment, the stone experiment, and the face-place-object localizer, which were always performed in this order.

Visual stimuli

Sixty-four scene images from four natural scene categories were selected according to three scene properties (Fig. 1A): entry-level category (beaches, city streets, forests, and train stations); temperature of the depicted scene (warm vs cold), and sound level (noisy vs quiet). The selection procedure ensured that all three properties vary independently within the image set.

Figure 1.
  • Download figure
  • Open in new tab
  • Download powerpoint
Figure 1.

A, Examples of scene images from four scene categories labeled as warm or cold (temperature property) and as noisy or quiet (sound-level property). B, Behavioral ratings of temperature and sound-level properties. C, Illustration of the image presentation in the main experiment. D, E, Image presentation in the temperature judgment condition (D) and in the sound-level judgment condition (E).

We measured the properties of scene images with a set of three behavioral experiments on Amazon MTurk. Workers in the behavioral experiments viewed each image and rated the temperature and the sound level on a 5-point Likert scale. Each type of property was measured in a separate block. The order of the tests was counterbalanced across participants. Before the rating, a detailed description of each scale was provided (the descriptions are reported at: https://github.com/yaelanjung/scenes_fmri), followed by six practice trials separately for the temperature rating and the sound-level rating.

In the first rating experiment, 27 MTurk workers rated 292 scene images, which were collected from the Internet. Ratings were z scored for each participant and averaged across participants for each image. We excluded 87 images with extreme temperature or sound scores, added 100 new images, and repeated the rating experiment with 15 participants, first on 305 images and then on a subset of 128 images. After this procedure, we selected the final set of 64 images and repeated the rating experiment on these images with 45 participants. These final ratings show that neither temperature nor sound level are related to category labels. A Kruskal–Wallis test was conducted to find the effect of category labels on the temperature rating, showing that there is no effect of category labels (Kruskal–Wallis test: χ2 = 3.54, p = 0.32; a Kruskal–Wallis test to measure the effect of category labels on sound-level ratings also reveal no effect of category labels: Kruskal–Wallis test: χ2 = 4.56, p = 0.21). Furthermore, the temperature ratings and the sound-level ratings were not correlated to each other (r = 0.037, p = 0.781; Fig. 1B). This final set of 64 images was used in the main experiment. The entire stimuli set and the rating data are available in a GitHub repository (https://github.com/yaelanjung/scenes_fmri).

Auditory stimuli

Ten different sound clips from different places (e.g., beaches, forests, and city streets) were downloaded from an open database, Freesound (freesound.org). The sounds were cut to a length of 12 s and were presented using MR-compatible, in-ear, noise-canceling headphones (model S14, Sensimetrics). The noisy and quiet sounds differed in two aspects. First, the noisy sounds had more auditory objects (i.e., birds, people, waves at a beach) than the quiet sounds. Also, based on the perceived loudness, which was estimated using a sound meter over the headphones, the noisy sounds had greater volume (on average, 81 dB) than the quiet sounds (on average, 53 dB; see Fig. 5A).

Stone stimuli

Six massage stones (∼6 × 8 × 2 cm) were used in the temperature experiment. To deliver stones at a particular temperature, three stones were soaked in warm water (∼45°C) in a plastic thermal cup, and the other three stones were soaked in cold water (∼9°C) for ∼10 min before the experiment. We measured the temperature of the stones before and after each run with a digital infrared thermometer. On average across different runs and participants, the cold stones were 11.0°C (SD = 1.3°C) before the runs and 15.3°C (SD = 0.93°C) after the runs. The warm stones were 44.4°C (SD = 0.93°C) before and 39.8°C (SD = 0.81°C) after the runs (see Fig. 4B).

Procedure

Main experiment

The main fMRI experiment was event related with six runs of image presentations with three different tasks (two runs for each task). In all three task conditions, participants viewed the same set of 64 images one at a time for 1 s, with a 5 s intertrial interval (ITI; Fig. 1C). The trial sequence also contained 16 blank trials with only the fixation cross visible; the first and last two trials in each run were always blank, and the other 12 blank trials were randomly distributed throughout the run. Visual stimuli were presented on the screen at a distance of ∼134 cm from the participant with images subtending ∼12.3° of visual angle. Stimulus presentation, the timing of all stimuli, and the recording of responses were controlled with custom MATLAB scripts, using the Psychophysics Toolbox (Brainard, 1997).

In the fixation monitoring runs, participants were asked to press a button to report whether the fixation cross was white (75% of the time) or red (25% of the time). In the temperature judgment runs, participants were asked to indicate whether the depicted scene looked warm or cold (Fig. 1D). In the sound-level judgment runs, participants were asked to indicate whether the depicted scene looked noisy or quiet (Fig. 1E). It was emphasized that there is no right or wrong answer in the temperature and sound-level judgment tasks. The same set of 64 images was repeated in all six runs, and the order of images was randomized in each run. In all three conditions, the response labels were presented during the entire run, and the color of the response label changed to green from black when a response was given and remained green until the end of the ITI (Fig. 1C).

Before the onset of each run, participants were reminded of the task by a text (e.g., temperature judgment run) on the screen and verbal instructions from the experimenter (behavioral data reported at: https://github.com/yaelanjung/scenes_fmri). Once participants verbally confirmed that they understood the task instruction for a given run, we tested the response buttons. Participants were asked to press each button one at a time (e.g., “If the picture looks warm, please press the 'WARM' button. Press the 'WARM' button to continue.”) to ensure that they were aware of the key instruction. All participants first performed the two fixation-monitoring runs and then the temperature runs and the sound-level runs interleaved (either T-S-T-S or S-T-S-T). The order was counterbalanced across participants.

Localizer experiments

The localizer experiment consisted of the sound experiment, the temperature experiment, and the face-place-object localizer. All of these experiments used a block design.

The stone experiment consisted of four runs and was performed by two experimenters, one in the control room and one in the examination room next to the MRI scanner. Each run had six blocks, and each block was one stone event (see Fig. 4A). At every onset or offset of a block, the experimenter in the control room signaled the event by presenting a colored flag (red, for warm stone; blue, for cold stone; green, for taking the stone away) through the control room window. The experimenter in the examination room dried a warm or cold stone on a towel and handed it to the left hand of the participant. Participants were holding the stone without any particular task for 10 s, and then the experimenter in the examination room took the stone from the participant's hand at the green flag signal.

The sound experiment consisted of two runs. Each run had 10 blocks, during which 12 s sound clips were played. The clips in half of the blocks were noisy sounds, and the other half of the blocks contained quiet sounds (see Fig. 5A). The order of the sound sequences was randomized for each run and for each participant. A fixation cross was presented at the center of the screen during the entire experiment.

The face-place-object localizer experiment consisted of two runs. Each run had four blocks, and there were four miniblocks with images of faces, objects, places, and scrambled object pictures in each block (Epstein and Kanwisher, 1998). Each miniblock was 16 s long, and each image was presented for 1 s with an intertrial interval of 500 ms. There were 12 s fixation periods between blocks and also at the beginning and the end of the run.

fMRI data acquisition

Imaging data were recorded on a 3 tesla Siemens Prisma MRI scanner with a 32-channel head coil at the Toronto Neuroimaging Center at the University of Toronto. High-resolution anatomic images were acquired with a MPRAGE (magnetization-prepared rapid acquisition gradient echo) protocol with multiband factor of 2. Images were then reconstructed using GRAPPA (generalized autocalibrating partial parallel acquisition; Griswold et al., 2002), with sagittal slices covering the whole brain (TI = 912 ms; TR = 1900 ms; TE = 2.67 ms; flip angle = 9°; voxel size = 1 × 1 × 1 mm; matrix size = 224 × 256 × 160 mm). Functional images for the main and localizer experiments were recorded with a multiband acquisition sequence (TR = 2 s; TE = 32 ms; flip angle = 70°; multiband factor = 4; voxel size = 2 × 2 × 2 mm; matrix size = 220 × 220 × 136 mm; 68 axial slices).

Data analyses

fMRI data preprocessing

fMRI data from the main experiment runs were motion corrected to one EPI image (the middle volume in the middle run), followed by spatial smoothing with a Gaussian kernel with 2 mm full-width at half-maximum (FWHM). Data were converted to the percentage signal change with respect to the mean of each run. Preprocessing was performed using AFNI (Cox, 1996).

Region of interest

PPA was defined using data from the face-place-object localizer runs. fMRI data from this localizer were preprocessed the same way as the main experiment data, but spatially smoothed with a 4 mm FWHM Gaussian filter. Data were further processed using a general linear model (GLM; 3dDeconvolve in AFNI) with regressors for all four image types (scenes, faces, objects, scrambled objects). PPA was defined as contiguous clusters of voxels with significant contrast [q < 0.05, corrected using false discovery rate (FDR)] of scenes > (faces and objects) (Epstein and Kanwisher, 1998).

The representation of the left hand in somatosensory cortex (S1) was also functionally defined using data from the temperature localizer runs as we are interested in a region that specifically corresponds to thermal perception driven by the stone stimuli. fMRI data from the temperature localizer runs were preprocessed the same way as the main experiment data and spatially smoothed with a 4 mm FWHM Gaussian filter. Data were then further processed using a general linear model (3dDeconvolve in AFNI) with regressors for stone events (stone and no stone), regardless of temperature. S1 for the left hand was defined as a contiguous cluster of voxels in the right hemisphere with significant contrast (q < 0.05, corrected using false discovery rate) of stone > no stone. In all participants except one, we observed a cluster with significant contrast near the central gyrus in the right hemisphere (contralateral to the left hand where the stone was given). The cluster was then masked by the anatomic extent of the post-central gyrus, which is known as the location of the primary somatosensory cortex. The subject for whom we could not functionally define S1 was excluded from the decoding analysis performed in this area.

The following anatomically defined regions of interest (ROIs) were extracted using a probabilistic atlas in AFNI (DD_Desai_MPM; Destrieux et al., 2010): middle frontal gyrus (MFG); superior frontal gyrus (SFG); and inferior frontal gyrus (IFG). Anatomical masks for primary auditory cortex (Heschl's gyrus) were made available by Norman-Haignere et al. (2013). After nonlinear alignment of each participant's anatomic image to Montreal Neurological Institute (MNI) space using the AFNI 3dQwarp function, the inverse of the alignment was used to project anatomic ROI masks back into original subject space using 3dNwarpApply. All decoding analyses, including for the anatomically defined ROIs, were performed in original subject space.

Decoding analyses

Multivoxel pattern analysis of the main experiment.

Decoding analysis was performed for each task condition: the temperature judgment condition and the sound-level judgment condition. Data from the fixation-monitoring condition were included as training but not as test data for both conditions. For each participant, we trained a linear support vector machine [SVM; using LIBSVM (Chang and Lin, 2011)] to assign the correct labels to the voxel activations inside an ROI based on the fMRI data from three runs: one of the task-specific runs as well both of the fixation-monitoring runs. The SVM decoder then produced predictions for the labels of the data in the left-out task run. This cross-validation procedure was repeated with the other task run left out, thus producing predicted labels for both task-specific runs. This procedure was performed for scene categories (beaches, forests, city streets, train stations), scene temperature (warm, cold), and sound level (noisy, quiet) separately for temperature judgment and sound-level judgment runs. Decoding accuracy was assessed as the fraction of predictions with correct labels. Group-level statistics were computed over all 20 participants using one-tailed t tests, determining whether decoding accuracy was significantly above chance level (0.5 for temperature and sound-level; 0.25 for scene categories). In all the statistical tests performed here, we adjusted the significance of the t test for multiple comparisons using FDR (Westfall and Young, 1993).

To curb overfitting of the classifier to the training data, we reduced the dimensionality of the neural data by selecting a subset of voxels in each ROI. Voxel selection was performed by ranking voxels in the training data according to the F statistics of a one-way ANOVA of the activity of each voxel with temperature, sound level, or scene categories as the main factor (Pereira et al., 2009). We determined the optimal number of voxels by performing a nested leave-one-run-out (LORO) cross-validation within the training data. It should be noted that the voxel selection is performed only using the training data (nested cross-validation) to avoid double dipping. Once an optimal number of voxels was determined by voxel selection, a classifier was trained again using the entire training data (but with selected voxels) and tested on the test data. The test data were not used to select voxels or to train the classifier for a given cross-validation fold. Optimal voxel numbers varied by ROI and participant, showing an overall average of 99 voxels across all ROIs and participants in the analysis of the main experiment.

Multivariate pattern analysis of the localizer experiments.

To decode temperature from the localizer experiment, we performed a GLM analysis on the preprocessed data with regressors for warm and hot blocks as well as nuisance regressors for motion and scanner drift within the run. We obtained eight sets of β weights (four runs x two conditions) and performed a LORO cross-validation analysis using an SVM classifier. A voxel selection procedure (as described above) was also included in this decoding procedure. For each cross-validation fold, the voxels were ranked based on the F statistics of an ANOVA of the activity of each voxel with temperature as the factor. We determined the optimal number of voxels by performing a LORO cross-validation analysis within the training data (nested cross-validation; mean = 211.11).

There were only two runs of data in the sound experiment. Thus, we performed a GLM analysis at the level of blocks instead of runs. We used the sets of β weights for each block as input to the decoder. Specifically, in this GLM analysis, the preprocessed data from the sound localizer runs were analyzed with regressors for noisy and quiet sound blocks as well as nuisance regressors for motion and scanner drift within the run (second-order polynomial). As a result, we obtained 20 sets of β weights (10 blocks x 2 conditions). These sets of β values were used in a one-block-leave-out cross-validation analysis (similar to LORO, but on each block instead of runs) using an SVM. A voxel selection procedure (as described above) was also performed for each fold in the decoding procedure, and the average optimal number of voxels (across all ROIs) was 304.02. The decoding accuracy for each ROI and each subject was determined by averaging the performance across all the folds of the cross-validation procedure.

Cross-decoding analysis.

In the cross-decoding analysis, we trained a classifier on data from the localizer session and tested it with data from the main experiment. Since training and test data were completely separate in these conditions, we used all data from the localizer session for training and voxel selection. We here used the accuracy values from the whole-brain searchlight analysis to rank voxels (see the following section for the detail). Using the training data, we performed LORO cross-validation analyses with the number of selected voxels varying from 125 to 500 (step size, 25). We included voxels according to decreasing rank order of their decoding accuracies in the searchlight analysis. We compared the decoding performance for different voxel numbers and determined the optimal number of voxels (mean = 171.25), which showed the best decoding performance within the training data. The decoder was then trained on the entire training set, using the optimal number of voxels, and then tested on the data from the main experiment at those same voxel locations.

Searchlight analysis.

To explore representations of scene categories outside of predefined ROIs, we performed a searchlight analysis. We defined a cubic “searchlight” of 5 × 5 × 5 voxels (10 × 10 × 10 mm). The searchlight was centered on each voxel in turn (Kriegeskorte et al., 2006), and LORO cross-validation analysis was performed within each searchlight location using a linear SVM classifier (CosmoMVPA Toolbox; Oosterhof et al., 2016). Decoding accuracy at a given searchlight location was assigned to the central voxel.

For group analysis, we first coregistered each participant's anatomic brain to the MNI 152 template using a diffeomorphic transformation as calculated by AFNI 3dQWarp. We then used the same transformation parameters to register individual decoding accuracy maps to MNI space using 3dNWarpApply, followed by spatial smoothing with a 4 mm FWHM Gaussian filter. To identify voxels with decodable categorical information at the group level, we performed one-tailed t tests to test whether decoding accuracy at each searchlight location was above chance (nonvisual properties and cross-decoding, 0.5; categories, 0.25). After thresholding at p < 0.05 (one-tailed) from the t test, we conducted a cluster-level correction for multiple comparisons. We used AFNI 3dClustSim to conduct α probability simulations for each participant. The estimated smoothness parameters computed by 3dFWHMx were used to conduct the cluster simulation. In the simulations, a p value of 0.05 was used to threshold the simulated data before clustering, and a corrected α of 0.05 was used to determine the minimum cluster size. We used the average of the minimum cluster sizes across all the participants as the cluster threshold, which was 215 voxels.

Results

Decoding of scene attributes

To assess neural representations of temperature and sound level from images, we performed the multivariate pattern analysis (MVPA) for each ROI and for each task condition.

As illustrated in Figure 2A, we were able to decode temperature (warm vs cold) in PPA (mean decoding accuracy = 58.75%, t(19) = 2.67, q = 0.01) and in MFG (mean decoding accuracy = 62.5%, t(19) = 2.94, q = 0.02) when participants performed the temperature judgment task. However, when participants were performing the sound-level judgment task, the temperature property was decoded neither in PPA (mean accuracy = 50%, t(19) = 0, q = 0.62) nor in the MFG (mean accuracy = 50%, t(19) = 0, q = 0.62; Table 1, for decoding performance in other ROIs). Sound level (noisy vs quiet) was decoded in the sound-level judgment condition, in the MFG (mean accuracy = 62.5%, t(19) = 2.52, q = 0.01) and the IFG (mean accuracy = 63.75%, t(19) = 2.98, q = 0.02). However, when the participants were performing the temperature judgment task, sound level could not be decoded in these areas, in MFG (mean accuracy = 55%, t(19) = 0.94, q = 0.18) and in IFG (mean accuracy = 48.75%, t(19) = −0.29, q = 0.61). The statistical analyses in all ROIs are reported in Table 1.

Figure 2.
  • Download figure
  • Open in new tab
  • Download powerpoint
Figure 2.

Decoding from the main experiment. A−C, Decoding of temperature (A), sound-level properties (B), and categories (C) of scenes in each ROI in the temperature judgment condition (orange bars) and in the sound-level judgment condition (green bars). D, Confusion matrices from category decoding: each row indicates the presented category (in order of beach, city, forest, and station), and each column indicates the predicted category. Diagonal boxes (with a pink frame) indicate the percentage of corrected predictions (rounded to the nearest integer). Dots in each figure represent individual data points. Error bars indicate SEM. Significance of the one-sample t-tests (one-tailed) was adjusted for multiple comparisons using FDR. *q < 0.05, ***q < 0.001.

View this table:
  • View inline
  • View popup
Table 1.

Mean of decoding accuracy and statistical analysis in each ROI for the temperature and the sound-level properties

To directly examine the effect of task context on decoding of scene attributes, we performed a two-way ANOVA for each ROI with task (temperature judgment vs sound-level judgment) and decoded property (temperature vs sound level) as a factor. In all ROIs, there was no main effect of either task or decoded property (all F values, <3.32; p values, >0.07), as well as no significant interaction between the two (all F values, <3.3; p values, >0.07) except in MFG. In the MFG, although there was no main effect of task, or decoded property, the interaction between decoded property and task was significant (F(1,76) = 4.57, p = 0.003). These results indicate that task context impacts which type of nonvisual properties are represented in the MFG.

For scene temperature, it is possible that the color of the images may hold cues to the temperature. We explored this possibility by quantifying the contribution of color information to the temperature-specific fMRI data. We calculated the average L* (lightness), a* (red/green value), b* (blue/yellow value) values from the CIELAB color channels for individual images (Fig. 3A, examples) and regressed out all of these values from the fMRI data before we performed the GLM analysis (Fig. 3A, right). Using a multivariate regression where the behavioral temperature rating is predicted by the color information, we observe that the color information from the CIELAB channels for individual images is indeed related to the behavioral ratings for temperature (R2 = 0.40; Fig. 3B). When we performed the MVPA on the color-controlled β values from the GLM analysis (Fig. 3C), we found that decoding accuracy was impacted in the PPA (mean accuracy = 53.75%, t(19) = 1.14, p = 0.1), but in the MFG, temperature could be still decoded above chance (mean accuracy = 58.75%, t(19) = 2.10, p = 0.02). Thus, it appears that color information is a strong cue for scene temperature in the PPA, but less so in the MFG.

Figure 3.
  • Download figure
  • Open in new tab
  • Download powerpoint
Figure 3.

A, Left, Examples of the scene images and their CIELAB values. Right, Illustration of color-control regression analysis. B, Predicted temperature rating (from all three color channel values) plotted over the average temperature rating. Dots in the figure indicate the individual images used in the experiment. C, Decoding accuracies of temperature from the color-controlled analysis. Error bars indicate SEM. Error bars indicate SEM. Significance of the one-sample t-tests (one-tailed) was adjusted for multiple comparisons using FDR. #p (uncorrected) < 0.05.

Decoding of scene categories

Scene categories were decoded with a linear SVM classifier using the same data from each task condition but with category labels as ground truth. Here, chance level was 25% since there were four different scene categories.

As illustrated in Figure 2C, scene categories were decoded in both task conditions in PFC as well as in PPA. In the temperature judgment condition, decoding accuracies for scene categories were significantly higher than chance in PPA (mean accuracy = 42.50%, t(19) = 5.98, q < 0.001), in MFG (mean accuracy = 33.75%, t(19) = 2.33, q = 0.01), and in IFG (mean accuracy = 34.38%, t(19) = 3.68, q = 0.02). In the sound-level judgment condition, decoding accuracy for scene categories was significantly higher than chance in PPA (mean accuracy = 43.14%, t(19) = 5.78, q < 0.001), in IFG (mean accuracy = 31.25%, t(19) = 2.03, q = 0.04), in MFG (mean accuracy = 33.75%, t(19) = 2.33, q = 0.03) and in SFG (mean accuracy = 34.38%, t(19) = 2.88, q = 0.01; Table 2, other ROIs). The confusion matrices from decoding performance in each ROI show that the classifiers successfully decode all four categories [Fig. 2D; more confusion matrices are reported in a GitHub repository (http://github.com/yaelanjung/scenes_fmri]. To test whether the task impacts the decoding accuracy of scene categories, one-way ANOVA was performed on decoding accuracy with task as a factor in each ROI. In all ROIs, there was no significant effect of task on decoding accuracy for scene categories (all F values, <1.55; all p values, >0.22. Thus, these findings indicate that, unlike scene attributes, scene categories are represented in PFC regions regardless of task context.

View this table:
  • View inline
  • View popup
Table 2.

Mean of decoding accuracy and statistical analysis in each ROI for scene category

As we identified the neural representations of temperature and sound level in scene images, we here further explore these representations to better understand how scene content is represented in PFC. Previous research has demonstrated that the prefrontal cortex represents scene categories at an abstract level, which generalizes between visual and auditory input (Jung et al., 2018). Thus, it is possible that the prefrontal cortex represents scene attributes, temperature, and sound-level information at such an abstract level as well. To examine this possibility, we first tested how both properties are represented in the brain when participants perceived them through direct sensory stimulation. We then performed cross-decoding analyses from direct sensations to percepts inferred from images to explore whether scene attributes are also represented beyond specific sensory domains.

Decoding of direct percepts and cross-decoding

Temperature properties

We first examined the neural representation of temperature that arises from direct thermal sensation. In the localizer experiment, participants were holding in their left hand a stone that was either warm (∼42°C, averaged across before and after each run) or cold (∼13°C, averaged across before and after each run; Fig. 4A,B). We performed a LORO cross-validation analysis on these data to decode the temperature across different brain regions. In this analysis, we included the representation of the left hand in somatosensory cortex. Since the stone stimulus was always handed to the left hand of participants, we only considered the contralateral (right) side of somatosensory cortex. We also included the prefrontal regions, inferior, middle, and superior frontal gyri, in which we observed neural coding of inferred thermal sensation from the main experiment.

Figure 4.
  • Download figure
  • Open in new tab
  • Download powerpoint
Figure 4.

Decoding of temperature property. A, Illustration of the stimuli presentation and the experiment design for the temperature localizer runs. B, Temperature of warm and cold stones before and after runs. C, Decoding of temperature from thermal sensations. D, Schematic of cross-decoding analysis across direct and indirect thermal percepts. E, Decoding accuracy of the cross-decoding analysis in the post-central sulcus and MFG. Dots in each figure represent individual data points. Error bars indicate SEM. Significance of the one-sample t-tests (one-tailed) was adjusted for multiple comparisons using FDR. *q < 0.05, #p (uncorrected) < 0.05.

As illustrated in Figure 4C, the decoder was significantly more accurate than chance (50%) in the right S1 (mean accuracy = 57.89%, t(18) = 2.584, q = 0.031), suggesting that the temperature of the stones is represented in S1. Furthermore, the classifier can predict the temperature of the stones in the MFG (mean accuracy = 60.62%, t(19) = 2.43, q = 0.044), indicating that, just like the inferred temperature information from scene images, the thermal percept is also represented in PFC.

Having identified brain regions that represent temperature delivered from direct sensations, we here investigate whether such neural representation is similar to the percepts inferred from images, which we observed in the main experiment. We do so by performing a cross-decoding analysis in the ROIs that showed significant decoding of sensory properties from images as well as direct percepts. For the temperature property, only MFG shows significant decoding of temperature from images (Fig. 2A) and from direct percepts (Fig. 4C). An SVM classifier was trained on the data from the direct percept of temperature in the localizer experiment and then tested on the data from viewing scene images in the main experiment (Fig. 4D).

Cross-decoding from the stone experiment to the main scene image experiment was significantly above chance in MFG (mean accuracy = 55%, t(19) = 1.9, p = 0.036), suggesting that the representation of temperature generalizes well from direct to inferred percepts in this brain region (Fig. 4E).

Sound-level properties

Similarly, we examined whether direct auditory sensations generalize to the neural representations of inferred sounds that were identified in the main experiment. First, we tested neural representations of noisy and quiet sounds from actual auditory percepts; an SVM classifier was trained and tested on the neural activity patterns from the sound localizer runs using LORO cross-validation. As illustrated in Figure 5B, the decoder correctly classified the sound level of scene sounds in primary auditory cortex (A1; mean accuracy = 84%, t(19) = 9.95, q < 0.001) and in the IFG (mean accuracy = 58%, t(19) = 3.046, q = 0.006).

Figure 5.
  • Download figure
  • Open in new tab
  • Download powerpoint
Figure 5.

Decoding of sound properties. A, Illustration of the design of the sound experiment, and amplitude of an exemplar of the noisy sound clips and an exemplar of the quiet sound clips used in the localizer experiment. B, Sound-level decoding accuracy in each ROI. C, Schematic of cross-decoding analysis across direct and indirect auditory percepts. D. Decoding performance of the cross-decoding analysis in IFG. Dots in each figure represent individual data points. Error bars indicate SEM. Significance of the one-sample t-tests (one-tailed) was adjusted for multiple comparisons using FDR. *q < 0.05, ***q < 0.001, #p (uncorrected) < 0.05.

As we confirmed that there are distinct patterns of neural representation for each type of sound level in both primary auditory cortex and PFC, we further examined whether these neural codes generalize to the sound level inferred from images. Cross-decoding of sound level was tested in IFG, which was the only brain area where we could decode both inferred (Fig. 2B) and direct (Fig. 5B) sensations of sound level. In IFG, decoding succeeded at mean accuracy = 56.25% (t(19) = 2.032, p = 0.028; Fig. 5D).

These results indicate that in PFC, scene attributes are represented at an abstract level and generalize across sensory modalities, from direct sensations of temperature and sound level to scene properties inferred from visual percepts.

Searchlight analysis

To explore how scene information is represented beyond the predefined ROIs, we performed a whole-brain searchlight analysis with a 5 × 5 × 5 voxels (10 × 10 × 10 mm) cube searchlight. The same decoding analyses for scene attributes (temperature and sound level) and scene categories as in the ROI-based analysis were performed at each searchlight location using a linear SVM classifier, followed by a cluster-level correction for multiple comparisons.

Confirming the ROI-based analysis, both temperature and scene categories were decoded from PFC in the temperature judgment condition (Fig. 6A). Clusters for category decoding were located in the anterior part of the MFG and SFG (Fig. 6A, purple), whereas the clusters for temperature decoding were located in posterior and medial parts of PFC (Fig. 6A, orange) with little overlap.

Figure 6.
  • Download figure
  • Open in new tab
  • Download powerpoint
Figure 6.

A, B, Searchlight maps of temperature and category decoding in the temperature judgment condition (A) and sound-level and category decoding in the sound-level judgment condition (B). Maps are thresholded at p < 0.05 and cluster-wise correction was conducted for multiple comparisons (minimal cluster size, 215). MNI coordinates: A: x = 20, y = −10, z = 15; B, x = −5, y = 32, z = 26.

In the sound-level judgment condition, scene categories were decoded in the anterior and posterior parts of PFC, whereas the sound-level properties were decoded mainly in the medial prefrontal gyrus. A small overlap between the clusters can be observed in the MFG (Fig. 6B).

Cross-decoding from the localizer experiment (direct percepts) to the main experiment (indirect percepts) was also performed in the whole brain using a searchlight analysis (Fig. 7). Consistent with the ROI-based analysis, we identified clusters showing successful cross-decoding of temperature and sound level in the prefrontal regions. Specifically, for temperature decoding, we found three clusters overlapping with SFG in both hemispheres as well as with the right MFG and the left IFG. For sound-level decoding, a cluster with successful decoding accuracy (p < 0.05) was located in the right MFG and IFG. We did not find any other clusters anywhere else in the brain that allowed for cross-modal decoding.

Figure 7.
  • Download figure
  • Open in new tab
  • Download powerpoint
Figure 7.

Searchlight maps of cross-decoding analysis for temperature (A) and sound-level (B) decoding. Maps are thresholded at p < 0.05, and a cluster-wise correction was conducted for multiple comparisons (minimal cluster size, 215).

Discussion

The present study demonstrates that both scene categories and scene attributes are represented in PFC. Specifically, scene attributes are represented in PFC only when the information is task relevant. Across the two task conditions, the temperature of scenes was decoded only when participants were required to judge the temperature of a depicted scene, and the sound level was decoded only when participants were required to judge how noisy a scene was. Intriguingly, when scene properties were represented in PFC, they were computed at an abstract level, which transcended the level of specific sensory instances. Direct thermal or auditory percepts elicited neural activity patterns similar to those that were evoked when participants inferred these properties from scene images, suggesting that PFC generalizes across direct and indirect (inferred) perceptions.

The current study provides strong and clear evidence that PFC represents scene information, suggesting that scene information processing is not restricted to the scene-selective regions in the visual cortex but extends to the associative cortex. Our findings demonstrate that PFC represents both temperature and sound-level properties as well as scene category information. These findings are in line with those of previous studies showing that various aspects of scene information are processed beyond the scene-processing network in the visual cortex (Groen et al., 2018; Jung et al., 2018; Epstein and Baker, 2019).

The findings of the current study are also in line with previous work showing that neural representations in PFC can change according to task context (Harel et al., 2014; McKee et al., 2014; Bugatus et al., 2017). Both temperature and sound-level information were represented in PFC only when participants were making a decision about these properties; temperature properties were decoded in MFG when participants were judging scene temperature, but not when they were engaged in the sound-level judgment task. Similarly, sound-level properties of scenes were decoded in IFG and MFG only in the sound-level judgment condition. To some extent, these findings suggest that PFC might represent scene content that is not represented in the visual or auditory cortex; sound-level information of the scene was not represented in PPA or A1, but in PFC. Nonetheless, PFC represents this information when it is relevant to behavioral demands. These findings suggest that scene representations in PFC are largely modulated by task context or behavioral goals.

In contrast to scene attributes, the current study shows that scene categories are represented in both PPA and PFC regardless of task context. The decoding accuracy of scene categories does not vary systemically across different task conditions, suggesting that category information is represented in PFC even when category information is not directly related to behavioral demands. These findings are in line with those of previous behavioral studies, which showed that scene categories are processed automatically and obligatorily (Greene and Fei-Fei, 2014). Furthermore, several neuroimaging studies have shown successful decoding of scene categories using a passive-viewing paradigm without any task (Walther et al., 2009, 2011; Park et al., 2011). However, in these fMRI studies, participants could have been deliberately tracking categories of scenes, although it was not explicitly required. In the present study, participants were asked to perform tasks that were orthogonal to scene categories, making it less likely that they paid close attention to scene categories. Still, this finding is limited by the use of only four scene categories in the current study. As there are hundreds of important scene categories (Tversky and Hemenway, 1983; Xiao et al., 2016), we cannot conclude that all of those categories are unaffected by task context. Future research is needed to further explore how tasks impact category representations in the brain.

Does the same set of voxels in the prefrontal cortex represent scene information, regardless of whether the scene information is about categories or scene attributes? Although our ROI-based analysis shows that the same regions (IFG and MFG) represent scene categories and scene attributes, the searchlight analysis indicates an interesting possibility that there may be different sets of clusters, each representing scene categories and scene attributes, respectively. Within the same anatomic ROIs, the clusters representing categories appear more anteriorly than the clusters representing scene attributes, which appear more posteriorly within the superior and middle prefrontal cortices (Fig. 6A, sagittal view). Further, in the sound judgment condition (Fig. 6B), scene category representations appear in the right IFG, and representations of scene attributes appear in left IFG. Thus, it is possible that different parts of the prefrontal cortex each represent scene categories and scene attributes, thereby potentially explaining the different degree to which task influences these scene representations.

Our findings might appear to contradict previous work showing that distinct neural signals for categories are diluted in PFC (Bracci et al., 2017; Bugatus et al., 2017). For instance, Bugatus et al. (2017) showed that the neural representations for visual categories, such as faces, words, and body parts, are evident in the ventral temporal cortex under different task contexts, but not in the prefrontal regions; PFC represents task context rather than categories of visual stimuli. There are, however, differences in the level of categories across the two studies (the current study; Bugatus et al., 2017) that could contribute to category-specific neural activity patterns in PFC. Unlike in the study by Bugatus et al. (2017), where superordinate categories, such as faces, pseudowords, and body parts were used, the current study uses basic-level scene categories, which are the “default” level of representing sensory input (Jolicoeur et al., 1984; Grill-Spector and Kanwisher, 2005; Mack et al., 2008). Several studies have demonstrated that such basic-level categories are accessible with very limited exposure (Mack et al., 2008; Mack and Palmeri, 2010) and with limited attention or cognitive resource (Li et al., 2002; Greene and Fei-Fei, 2014). Further, it should be noted that tasks are largely different across the two studies, and in the current study, the tasks in different conditions are still all related to scene understanding. Thus, future research is needed to better understand how different tasks may or may not impact category-selective representations in PFC.

Relatedly, Bracci et al. (2017) also show that neural representations of object categories are diluted by task context. PFC was found to represent object categories only when participants made judgments of semantic relationships among different objects (which is referred to as category task; Bracci et al., 2017, their Fig. 2). The seeming inconsistency with our study might be because of different definitions of categories. We define scene categories as basic-level category (Tversky and Hemenway, 1983) and used multiple exemplars from each category (i.e., multiple beach images). Bracci et al. (2017) used different object categories under the same superordinate category (i.e., piano, violin, drum, for instruments) and defined category information based on semantic relationships among the individual items, manifested in a continuous measure. Thus, although basic-level category information is preserved when the task is not related to category information (as found in the current study), richer information about the category, or semantic relationships among individual objects, might only be decodable from PFC when the task requires access to category information (as shown in Bracci et al., 2017). Exploring this intriguing possibility in future work will require a different experiment design than is used in the current study.

The current study also broadens our understanding of what kinds of scene content are represented in the PPA. To our knowledge, our findings are the first to show that temperature can be decoded in PPA, just like scene categories (Walther et al., 2009) and scene properties related to spatial structure of a scene (e.g., openness; Park and Chun, 2009). The representation of temperature in the PPA appears to be driven by color to a large extent, suggesting that scene representations in PPA primarily rely on visual features.

We here provide evidence supporting the notion that representations in PFC generalize across sensory modalities, such as between the perceptions of images and stones, or images and sounds. Specifically, our findings are novel in that generalization across sensory modalities can occur when sensory information is inferred, not directly perceived. Although previous studies have shown that PFC represents sensory experience regardless of sensory modalities, participants in these studies always experienced direct stimulations in the corresponding modalities, such as hearing beach sounds and looking at a beach image (Vetter et al., 2014; Jung et al., 2018). The present study extends these findings by demonstrating that percepts inferred from a different modality (i.e., temperature or sound cues inferred from images) can elicit similar neural codes as the direct sensation would.

To summarize, the current study demonstrates that both scene categories and scene attributes are represented in PFC; however, neural representations of scene attributes (temperature and sound level) in PFC, not scene categories, are modulated by task context. These findings suggest that PFC flexibly represents scene content that is relevant to behavioral demands, allowing for efficient adaptation of behavior to varying situations in the real world.

Footnotes

  • This work was supported by the Social Sciences and Humanities Research Council of Canada (Grant 430-2017-01189 to D.B.W.), and the Natural Sciences and Engineering Research Council of Canada (Grant RGPIN-2015-06696 to D.B.W.).

  • The authors declare no competing financial interests.

  • Correspondence should be addressed to Yaelan Jung at jung.yaelan{at}gmail.com

SfN exclusive license.

References

  1. ↵
    1. Bracci S,
    2. Daniels N,
    3. Op de Beeck H
    (2017) Task context overrules object-and category-related representational content in the human parietal cortex. Cereb Cortex 27:310–321.
    OpenUrl
  2. ↵
    1. Brainard DH
    (1997) The psychophysics toolbox. Spat Vis 10:433–436. pmid:9176952
    OpenUrlCrossRefPubMed
  3. ↵
    1. Bugatus L,
    2. Weiner KS,
    3. Grill-Spector K
    (2017) Task alters category representations in prefrontal but not high-level visual cortex. Neuroimage 155:437–449.
    OpenUrl
  4. ↵
    1. Chadick JZ,
    2. Gazzaley A
    (2011) Differential coupling of visual cortex with default or frontal-parietal network based on goals. Nat Neurosci 14:830–832. doi:10.1038/nn.2823 pmid:21623362
    OpenUrlCrossRefPubMed
  5. ↵
    1. Chang CC,
    2. Lin CJ
    (2011) LIBSVM: a library for support vector machines. ACM Trans Intell Syst Technol 2:1–27. doi:10.1145/1961189.1961199
    OpenUrlCrossRef
  6. ↵
    1. Cox RW
    (1996) AFNI: software for analysis and visualization of functional magnetic resonance neuroimages. Comput Biomed Res 29:162–173. doi:10.1006/cbmr.1996.0014 pmid:8812068
    OpenUrlCrossRefPubMed
  7. ↵
    1. Destrieux C,
    2. Fischl B,
    3. Dale A,
    4. Halgren E
    (2010) Automatic parcellation of human cortical gyri and sulci using standard anatomical nomenclature. Neuroimage 53:1–15. doi:10.1016/j.neuroimage.2010.06.010 pmid:20547229
    OpenUrlCrossRefPubMed
  8. ↵
    1. Dilks DD,
    2. Julian JB,
    3. Paunov AM,
    4. Kanwisher N
    (2013) The occipital place area is causally and selectively involved in scene perception. J Neurosci 33:1331–1336. doi:10.1523/JNEUROSCI.4081-12.2013 pmid:23345209
    OpenUrlAbstract/FREE Full Text
  9. ↵
    1. Epstein RA,
    2. Baker CI
    (2019) Scene perception in the human brain. Annu Rev Vis Sci 5:373–397. doi:10.1146/annurev-vision-091718-014809 pmid:31226012
    OpenUrlCrossRefPubMed
  10. ↵
    1. Epstein R,
    2. Kanwisher N
    (1998) A cortical representation of the local visual environment. Nature 392:598–601. doi:10.1038/33402 pmid:9560155
    OpenUrlCrossRefPubMed
  11. ↵
    1. Greene MR,
    2. Fei-Fei L
    (2014) Visual categorization is automatic and obligatory: evidence from Stroop-like paradigm. J Vis 14(1):14, 1–11. doi:10.1167/14.1.14
    OpenUrlAbstract/FREE Full Text
  12. ↵
    1. Greene MR,
    2. Oliva A
    (2009) Recognition of natural scenes from global properties: seeing the forest without representing the trees. Cogn Psychol 58:137–176. doi:10.1016/j.cogpsych.2008.06.001 pmid:18762289
    OpenUrlCrossRefPubMed
  13. ↵
    1. Greene MR,
    2. Baldassano C,
    3. Esteva A,
    4. Beck DM,
    5. Fei-Fei L
    (2016) Visual scenes are categorized by function. J Exp Psychol Gen 145:82–94. doi:10.1037/xge0000129
    OpenUrlCrossRefPubMed
  14. ↵
    1. Grill-Spector K,
    2. Kanwisher N
    (2005) Visual recognition: as soon as you know it is there, you know what it is. Psychol Sci 16:152–160. doi:10.1111/j.0956-7976.2005.00796.x
    OpenUrlCrossRefPubMed
  15. ↵
    1. Griswold MA,
    2. Jakob PM,
    3. Heidemann RM,
    4. Nittka M,
    5. Jellus V,
    6. Wang J,
    7. Kiefer B,
    8. Haase A
    (2002) Generalized autocalibrating partially parallel acquisitions (GRAPPA). Magn Reson Med 47:1202–1210. doi:10.1002/mrm.10171
    OpenUrlCrossRefPubMed
  16. ↵
    1. Groen II,
    2. Greene MR,
    3. Baldassano C,
    4. Fei-Fei L,
    5. Beck DM,
    6. Baker CI
    (2018) Distinct contributions of functional and deep neural network features to representational similarity of scenes in human brain and behavior. Elife 7:e32962. doi:10.7554/eLife.32962
    OpenUrlCrossRef
  17. ↵
    1. Harel A,
    2. Kravitz DJ,
    3. Baker CI
    (2014) Task context impacts visual object processing differentially across the cortex. Proc Natl Acad Sci U S A 111:E962–E971. doi:10.1073/pnas.1312567111
    OpenUrlAbstract/FREE Full Text
  18. ↵
    1. Jolicoeur P,
    2. Gluck MA,
    3. Kosslyn SM
    (1984) Pictures and names: making the connection. Cogn Psychol 16:243–275. doi:10.1016/0010-0285(84)90009-4
    OpenUrlCrossRefPubMed
  19. ↵
    1. Jung Y,
    2. Larsen B,
    3. Walther DB
    (2018) Modality-independent coding of scene categories in prefrontal cortex. J Neurosci 38:5969–5981. doi:10.1523/JNEUROSCI.0272-18.2018
    OpenUrlAbstract/FREE Full Text
  20. ↵
    1. Kravitz DJ,
    2. Peng CS,
    3. Baker CI
    (2011) Real-world scene representations in high-level visual cortex: it's the spaces more than the places. J Neurosci 31:7322–7333. doi:10.1523/JNEUROSCI.4588-10.2011
    OpenUrlAbstract/FREE Full Text
  21. ↵
    1. Kriegeskorte N,
    2. Goebel R,
    3. Bandettini P
    (2006) Information-based functional brain mapping. Proc Natl Acad Sci U S A 103:3863–3868. doi:10.1073/pnas.0600244103
    OpenUrlAbstract/FREE Full Text
  22. ↵
    1. Li FF,
    2. VanRullen R,
    3. Koch C,
    4. Perona P
    (2002) Rapid natural scene categorization in the near absence of attention. Proc Natl Acad Sci U S A 99:9596–9601. doi:10.1073/pnas.092277599
    OpenUrlAbstract/FREE Full Text
  23. ↵
    1. Mack ML,
    2. Palmeri TJ
    (2010) Decoupling object detection and categorization. J Exp Psychol Hum Percept Perform 36:1067–1079. doi:10.1037/a0020254 pmid:20731505
    OpenUrlCrossRefPubMed
  24. ↵
    1. Mack ML,
    2. Gauthier I,
    3. Sadr J,
    4. Palmeri TJ
    (2008) Object detection and basic-level categorization: sometimes you know it is there before you know what it is. Psychon Bull Rev 15:28–35. doi:10.3758/pbr.15.1.28 pmid:18605476
    OpenUrlCrossRefPubMed
  25. ↵
    1. Maguire E
    (2001) The retrosplenial contribution to human navigation: a review of lesion and neuroimaging findings. Scand J Psychol 42:225–238. doi:10.1111/1467-9450.00233 pmid:11501737
    OpenUrlCrossRefPubMed
  26. ↵
    1. McKee JL,
    2. Riesenhuber M,
    3. Miller EK,
    4. Freedman DJ
    (2014) Task dependence of visual and category representations in prefrontal and inferior temporal cortices. J Neurosci 34:16065–16075. doi:10.1523/JNEUROSCI.1660-14.2014 pmid:25429147
    OpenUrlAbstract/FREE Full Text
  27. ↵
    1. Norman-Haignere S,
    2. Kanwisher N,
    3. McDermott JH
    (2013) Cortical pitch regions in humans respond primarily to resolved harmonics and are located in specific tonotopic regions of anterior auditory cortex. J Neurosci 33:19451–19469. doi:10.1523/JNEUROSCI.2880-13.2013 pmid:24336712
    OpenUrlAbstract/FREE Full Text
  28. ↵
    1. Oosterhof NN,
    2. Connolly AC,
    3. Haxby JV
    (2016) CoSMoMVPA: multi-modal multivariate pattern analysis of neuroimaging data in Matlab/GNU Octave. Front Neuroinform 10:27. doi:10.3389/fninf.2016.00027 pmid:27499741
    OpenUrlCrossRefPubMed
  29. ↵
    1. Park S,
    2. Chun MM
    (2009) Different roles of the parahippocampal place area (PPA) and retrosplenial cortex (RSC) in panoramic scene perception. Neuroimage 47:1747–1756. doi:10.1016/j.neuroimage.2009.04.058
    OpenUrlCrossRefPubMed
  30. ↵
    1. Park S,
    2. Brady TF,
    3. Greene MR,
    4. Oliva A
    (2011) Disentangling scene content from spatial boundary: complementary roles for the parahippocampal place area and lateral occipital complex in representing real-world scenes. J Neurosci 31:1333–1340. doi:10.1523/JNEUROSCI.3885-10.2011 pmid:21273418
    OpenUrlAbstract/FREE Full Text
  31. ↵
    1. Pereira F,
    2. Mitchell T,
    3. Botvinick M
    (2009) Machine learning classifiers and fMRI: a tutorial overview. Neuroimage 45:S199–S209. doi:10.1016/j.neuroimage.2008.11.007 pmid:19070668
    OpenUrlCrossRefPubMed
  32. ↵
    1. Persichetti AS,
    2. Dilks DD
    (2019) Distinct representations of spatial and categorical relationships across human scene-selective cortex. Proc Natl Acad Sci U S A 116:21312–21317. doi:10.1073/pnas.1903057116
    OpenUrlAbstract/FREE Full Text
  33. ↵
    1. Silson EH,
    2. Steel AD,
    3. Baker CI
    (2016) Scene-selectivity and retinotopy in medial parietal cortex. Front Hum Neurosci 10:412. doi:10.3389/fnhum.2016.00412 pmid:27588001
    OpenUrlCrossRefPubMed
  34. ↵
    1. Silson EH,
    2. Gilmore AW,
    3. Kalinowski SE,
    4. Steel A,
    5. Kidder A,
    6. Martin A,
    7. Baker CI
    (2019) A posterior–anterior distinction between scene perception and scene construction in human medial parietal cortex. J Neurosci 39:705–717. doi:10.1523/JNEUROSCI.1219-18.2018 pmid:30504281
    OpenUrlAbstract/FREE Full Text
  35. ↵
    1. Tversky B,
    2. Hemenway K
    (1983) Categories of environmental scenes. Cogn Psychol 15:121–149. doi:10.1016/0010-0285(83)90006-3
    OpenUrlCrossRef
  36. ↵
    1. Vetter P,
    2. Smith FW,
    3. Muckli L
    (2014) Decoding sound and imagery content in early visual cortex. Curr Biol 24:1256–1262. doi:10.1016/j.cub.2014.04.020 pmid:24856208
    OpenUrlCrossRefPubMed
  37. ↵
    1. Walther DB,
    2. Caddigan E,
    3. Fei-Fei L,
    4. Beck DM
    (2009) Natural scene categories revealed in distributed patterns of activity in the human brain. J Neurosci 29:10573–10581. doi:10.1523/JNEUROSCI.0559-09.2009
    OpenUrlAbstract/FREE Full Text
  38. ↵
    1. Walther DB,
    2. Chai B,
    3. Caddigan E,
    4. Beck DM,
    5. Fei-Fei L
    (2011) Simple line drawings suffice for functional MRI decoding of natural scene categories. Proc Natl Acad Sci U S A 108:9661–9666. doi:10.1073/pnas.1015666108
    OpenUrlAbstract/FREE Full Text
  39. ↵
    1. Westfall PH,
    2. Young SS
    (1993) Resampling-based multiple testing: examples and methods for p-value adjustment. New York: Wiley.
  40. ↵
    1. Xiao J,
    2. Ehinger KA,
    3. Hays J,
    4. Torralba A,
    5. Oliva A
    (2016) Sun database: exploring a large collection of scene categories. Int J Comput Vis 119:3–22. doi:10.1007/s11263-014-0748-y
    OpenUrlCrossRef
  41. ↵
    1. Zanto TP,
    2. Rubens MT,
    3. Thangavel A,
    4. Gazzaley A
    (2011) Causal role of the prefrontal cortex in top-down modulation of visual processing and working memory. Nat Neurosci 14:656–661. doi:10.1038/nn.2773 pmid:21441920
    OpenUrlCrossRefPubMed
Back to top

In this issue

The Journal of Neuroscience: 41 (34)
Journal of Neuroscience
Vol. 41, Issue 34
25 Aug 2021
  • Table of Contents
  • Table of Contents (PDF)
  • About the Cover
  • Index by author
  • Ed Board (PDF)
Email

Thank you for sharing this Journal of Neuroscience article.

NOTE: We request your email address only to inform the recipient that it was you who recommended this article, and that it is not junk mail. We do not retain these email addresses.

Enter multiple addresses on separate lines or separate them with commas.
Neural Representations in the Prefrontal Cortex Are Task Dependent for Scene Attributes But Not for Scene Categories
(Your Name) has forwarded a page to you from Journal of Neuroscience
(Your Name) thought you would be interested in this article in Journal of Neuroscience.
CAPTCHA
This question is for testing whether or not you are a human visitor and to prevent automated spam submissions.
Print
View Full Page PDF
Citation Tools
Neural Representations in the Prefrontal Cortex Are Task Dependent for Scene Attributes But Not for Scene Categories
Yaelan Jung, Dirk B. Walther
Journal of Neuroscience 25 August 2021, 41 (34) 7234-7245; DOI: 10.1523/JNEUROSCI.2816-20.2021

Citation Manager Formats

  • BibTeX
  • Bookends
  • EasyBib
  • EndNote (tagged)
  • EndNote 8 (xml)
  • Medlars
  • Mendeley
  • Papers
  • RefWorks Tagged
  • Ref Manager
  • RIS
  • Zotero
Respond to this article
Request Permissions
Share
Neural Representations in the Prefrontal Cortex Are Task Dependent for Scene Attributes But Not for Scene Categories
Yaelan Jung, Dirk B. Walther
Journal of Neuroscience 25 August 2021, 41 (34) 7234-7245; DOI: 10.1523/JNEUROSCI.2816-20.2021
Twitter logo Facebook logo Mendeley logo
  • Tweet Widget
  • Facebook Like
  • Google Plus One

Jump to section

  • Article
    • Abstract
    • Introduction
    • Materials and Methods
    • Results
    • Discussion
    • Footnotes
    • References
  • Figures & Data
  • Info & Metrics
  • eLetters
  • PDF

Keywords

  • fMRI
  • MVPA
  • PFC
  • PPA
  • scene categorization
  • task dependency

Responses to this article

Respond to this article

Jump to comment:

No eLetters have been published for this article.

Related Articles

Cited By...

More in this TOC Section

Research Articles

  • Regional Excitatory–Inhibitory Balance Relates to Self-Reference Effect on Recollection via the Precuneus/Posterior Cingulate Cortex–Medial Prefrontal Cortex Connectivity
  • Attention Alters Population Spatial Frequency Tuning
  • Sequence Diversity and Encoded Enzymatic Differences of Monocistronic L1 ORF2 mRNA Variants in the Aged Normal and Alzheimer's Disease Brain
Show more Research Articles

Behavioral/Cognitive

  • Attention Alters Population Spatial Frequency Tuning
  • Complex Impact of Stimulus Envelope on Motor Synchronization to Sound
  • The Molecular Substrates of Second-Order Conditioned Fear in the Basolateral Amygdala Complex
Show more Behavioral/Cognitive
  • Home
  • Alerts
  • Follow SFN on BlueSky
  • Visit Society for Neuroscience on Facebook
  • Follow Society for Neuroscience on Twitter
  • Follow Society for Neuroscience on LinkedIn
  • Visit Society for Neuroscience on Youtube
  • Follow our RSS feeds

Content

  • Early Release
  • Current Issue
  • Issue Archive
  • Collections

Information

  • For Authors
  • For Advertisers
  • For the Media
  • For Subscribers

About

  • About the Journal
  • Editorial Board
  • Privacy Notice
  • Contact
  • Accessibility
(JNeurosci logo)
(SfN logo)

Copyright © 2025 by the Society for Neuroscience.
JNeurosci Online ISSN: 1529-2401

The ideas and opinions expressed in JNeurosci do not necessarily reflect those of SfN or the JNeurosci Editorial Board. Publication of an advertisement or other product mention in JNeurosci should not be construed as an endorsement of the manufacturer’s claims. SfN does not assume any responsibility for any injury and/or damage to persons or property arising from or related to any use of any material contained in JNeurosci.