Abstract
Accumulating evidence suggests that visual object understanding involves a rapid feedforward sweep, after which subsequent recurrent interactions are necessary. The extent to which recurrence plays a critical role in object processing remains to be determined. Recent studies have demonstrated that recurrent processing is modulated by increasing semantic demands. Differentially from previous studies, we used dynamic causal modeling to model neural activity recorded with magnetoencephalography while 14 healthy humans named two sets of visual objects that differed in the degree of semantic accessing demands, operationalized in terms of the values of basic psycholinguistic variables associated with the presented objects (age of acquisition, frequency, and familiarity). This approach allowed us to estimate the directionality of the causal interactions among brain regions and their associated connectivity strengths. Furthermore, to understand the dynamic nature of connectivity (i.e., the chronnectome; Calhoun et al., 2014) we explored the time-dependent changes of effective connectivity during a period (200–400 ms) where adding semantic-feature information improves modeling and classifying visual objects, at 50 ms increments. First, we observed a graded involvement of backward connections, that became active beyond 200 ms. Second, we found that semantic demands caused a suppressive effect in the backward connection from inferior frontal cortex (IFC) to occipitotemporal cortex over time. These results complement those from previous studies underscoring the role of IFC as a common source of top-down modulation, which drives recurrent interactions with more posterior regions during visual object recognition. Crucially, our study revealed the inhibitory modulation of this interaction in situations that place greater demands on the conceptual system.
- dynamic causal modeling
- effective connectivity
- recurrent interactions
- top-down modulation
- visual object naming
Introduction
Current conceptions of visual object processing consider that the ventral visual pathway (Ungerleider and Mishkin, 1982) is responsible for the transformation from visual image to meaningful object (Lupyan et al., 2010; Carlson et al., 2014; Mur, 2014). Anatomically, this route consists of an interconnected network of cortical regions moving rostrally from the occipitotemporal cortex (OTC) to the anterior-inferior temporal cortex, extending into the inferior frontal cortex (IFC; Kravitz et al., 2013). Information flow across this stream is supported by distinct associative white matter fibers. Although tractography cannot determine the direction of axonal projections in a tract (Acosta-Cabronero et al., 2011), studies of dynamic interactions showed that after an initial feedforward projection, recurrent processing constitutes a mechanism for the integration of higher-level semantic information with visual information (Bar et al., 2006; Chan et al., 2011; O'Reilly et al., 2013; Clarke et al., 2014; Mur and Kriegeskorte, 2014). The extent to which recurrent interactions critically contribute to object understanding is modulated by task-induced processing demands (Harel et al., 2014; Tang et al., 2014), and in particular by the specificity of semantic processing required by the task (Clarke et al., 2011, 2014; O'Reilly et al., 2013; Cichy et al., 2014; Harel et al., 2014). So that higher-order anterior areas are required for fine-grained differentiation processes (Moss et al., 2005; Clarke and Tyler, 2014; Clarke et al., 2014), which exert top-down influence upon posterior regions, instantiating recurrent interactions (Tang et al., 2014). Most of these studies have inferred top-down influences by indirect means (i.e., delays in the response latency between conditions). Here, we applied dynamic causal modeling (DCM) for evoked responses (David et al., 2006), to model the activity of brain regions recorded with magnetoencephalography (MEG) while participants named visual objects at the basic level, which were subsequently separated into a high- and low-demanding subsets according to values of psycholinguistic variables (i.e., familiarity, frequency, and age of acquisition) associated with the objects to be named (Krishnan et al., 2014). DCM allows us to understand how information propagates through brain regions (Kahan and Foltynie, 2013) by estimating the directionality of the causal interactions among brain regions and their associated connectivity strengths (Bianchi et al., 2013). The effects of psycholinguistic variables reflect differences in processing demands, which correlate with differences in the neurophysiologic processes underlying particular cognitive operations (Ellis et al., 2006; Graves et al., 2007). As the focus was on the role of recurrent connections, we compared functional architectures with and without backward connections (Garrido et al., 2007), and tested their contribution during visual naming of high- and low-semantically demanding items. Crucially, we explored the time-varying patterns of coupling among regions (the chronnectome; Calhoun et al., 2014), and estimated the effective connectivity during a period (beyond 200 ms) where adding semantic-feature information improves modeling and classifying visual objects (Clarke et al., 2014) at 50 ms increments. By comparing the models, we tested the contribution of feedforward and feedback interactions (Penny et al., 2010) at different moments in time (Calhoun et al., 2014).
Materials and Methods
Subjects
Fourteen adult subjects (9 female; mean age 36.64 years, SD = 8.42 years; mean formal education 16 years, SD = 1.84 years), without any history of neurological or psychiatric illness, volunteered for participation in the study, and gave written consent in accordance with the Declaration of Helsinki, after the nature of the procedures involved had been explained to them. Participants were right handed according to the Edinburgh Handedness Inventory (Oldfield, 1971). Of note, data from 10 of the participants were part of a previous study (Campo et al., 2013).
Naming task
Participants were required to name aloud black and white line drawings of common objects at the basic level (Clarke et al., 2011, 2014), thus ensuring that participants processed the objects at an individual level (Mur, 2014). The study used images from a visual confrontation naming task that was created in a previous study (Campo et al., 2013) by combining images from the Cambridge 64-item naming task (Bozeat et al., 2000) and the 175-item Philadelphia Naming Test (Roach et al., 1996). Images belonging to both tests were presented only once. From the pool of items, two subsets of 64 items of visual objects that differed in their semantic demands, operationalized in terms of the values of basic psycholinguistic variables associated with the objects to be named; specifically age of aquisition (AoA), frequency of occurrence, and concept familiarity (Krishnan et al., 2014 used a similar approach). These variables are significantly correlated among them, especially in large sets of pictorial stimuli (Graves et al., 2007; Wilson et al., 2009; Strijkers et al., 2010; Woollams, 2012). This fact, along with the consideration that their influence has been commonly accounted for by the degree of representation and connection in a distributed-representation network (Steyvers and Tenenbaum, 2005; Ellis et al., 2006; Patterson, 2007), led us to evaluate the modulation of connectivity using two different sets of objects showing a typical distribution across those psycholinguistic dimensions. Thus, items were sorted by concept familiarity, according to Spanish available norms (Sanfeliú and Fernández, 1996; Cuetos et al., 1999; Sebastián et al., 2000), and differed not only in terms of familiarity, but also on AoA, and frequency of occurrence (all p < 0.001). Concept familiarity was rated on a five-point scale from 1 = unfamiliar to 5 = highly familiar. Frequency was taken from the Alameda and Cuetos' (1995) dictionary of frequencies, which is based on a corpus of written texts comprising 2 million words (Cuetos et al., 1999). Both sets contain an equal number of living (32 exemplars) and nonliving objects (32 exemplars), which were also matched on picture complexity. As a measure of perceptual complexity we used JPEG size (Müller et al., 2008). Characteristics of the stimuli for both conditions are shown in Table 1. Conditions were labeled as “high-demanding” and “low-demanding” (Graves et al., 2007). The task was adapted for scanning purposes, such that in each trial participants first saw a fixation cross located centrally for 1000 ms, which was followed by a picture lasting in the display 1000 ms. Then, a question mark was shown indicating that participants have to overtly name the presented object as accurately as possible (Laaksonen et al., 2012). The next trial began when participants provided an answer or after 5000 ms elapsed. Picture naming performance was calculated in terms of the proportion of correct responses.
Characteristics of the stimuli comprising the 64 low-demanding and 64 high-demanding conditions
Data acquisition and analysis
Magnetoencephalography recordings.
MEG data were obtained using a whole-head 306 channel Vector-view system (Elekta-Neuromag), consisting of 102 magnetometers and 204 orthogonal planar gradiometers. The signal was recorded continuously at a sampling rate of 600 Hz with an online bandpass filter from 0.1 to 200 Hz. The head position relative to the sensor array was measured at the beginning of the session using four head position indicator coils. Before the recording session, the anatomical landmarks (nasion, and left and right periauricular) and extra points of the head shape were obtained using a 3D digitizer (Fastrak Polhemus). In addition, vertical electro-oculogram was recorded, with electrodes located supraorbitally and infraorbitally.
Static band channels were detected using the MaxFilter program (v2.2.10; Elektra-Neuromag), and were interpolated. The number of excluded channels varied between one and four (M = 2, SD = 1), which were found in nine participants. Artifacts were suppressed by applying a temporally extended signal–space separation method (Taulu and Hari, 2009), using a 10 s correlation window with a correlation limit of 0.9.
Magnetoencephalography data preprocessing and source localization.
Data were preprocessed and subsequently analyzed using Statistical Parametric Mapping (SPM8) academic software (Wellcome Trust Centre for Neuroimaging, UCL; http://www.fil.ion.ucl.ac.uk/spm/; Litvak et al., 2011 provides a detailed description) implemented in MATLAB (MathWorks). Data analyses were conducted using the 204 planar gradiometer channels. The continuous time series for each participant was processed with a Butterworth bandpass filter at 3–30 Hz and then were epoched off-line to obtain 500 ms data segments corresponding to a −100 to 400 ms peristimulus time. We analyzed epoched data during this period for each trial, for each condition (i.e., high-demanding and low-demanding), for each participant. Trials including eye blinks, or other myogenic or mechanical artifacts were removed using the thresholding criteria implemented in SPM8 [trials containing signal strength exceeding 3000 fT/cm (Furl et al., 2014)] and μ100 at electrooculogram channels were excluded). Epochs were baseline corrected from −100 to 0 ms, and then averaged.
The next step was to estimate the cortical origin of the neuronal response. For source reconstruction, a multiple sparse priors routine (as implemented in SPM8) was used (Friston et al., 2008), which uses 512 patches of activation that are iteratively reduced until an optimal number and location of active patches are found using a Bayesian greedy search. A 8196 vertex template cortical mesh in canonical Montreal Neurological Institute (MNI) anatomical space served as a brain model for the estimation of the current source distribution (Mattout et al., 2007). Coregistration to the MNI was done using the three anatomical landmarks, as well as the extra-digitalized points (i.e., headshape). This dipole mesh was used to calculate the forward solution using a digitalized single-shell model. The inverse solution was calculated over a time window from 0 to 400 ms after stimulus onset for each condition, and averaged over participants. As pointed by Strijkers et al. (2010), MEG studies of overt picture naming have shown that reliable measures of brain activity can be taken at least until 400 ms after picture onset (Hart et al., 1998; Maess et al., 2002; Clarke et al., 2011, 2013, 2014; Mousas et al., 2015). Furthermore, as two previous studies have estimated the onset of semantic processing of visual objects to be ∼200 ms (Hart et al., 1998; Clarke et al., 2014), we explored this time interval at 50 ms increments, from 200 to 400 ms. Source reconstructions were interpolated into MNI voxel space and analyzed using statistical parametric mapping (Kilner and Friston, 2010), as described by Moran et al. (2013). A contrast of high-demanding versus low-demanding based on the evoked related fields (ERF) was conducted at p < 0.005 (uncorrected) using a paired t test.
Effective connectivity analysis: dynamic causal modeling.
DCM is a hypothesis-driven method that relies on the specification of a plausible biophysical and physiological model of interacting brain regions (Stephan and Friston, 2007), and is therefore appropriate for situations where there is a priori knowledge and experimental control over the system of study (Cardin et al., 2011; Seghier et al., 2011). Bayesian model selection (BMS) is used to compare different models or to find the model with best evidence (Penny et al., 2004). This allows one to compare alternative hypotheses (models) of how measured data are caused (Friston, 2011). Thus, a critical factor is the architecture of the models, that is, the nodes and their interconnections (Rudrauf et al., 2008). As mentioned above, the selection of sources or nodes of the network architectures was based on inverse solutions (i.e., multiple sparse priors; Friston et al., 2008), therefore optimized for the particular subjects studied. Sources were chosen by the apparent signal propagation along the cerebral cortex observed in source reconstruction results (Rudrauf et al., 2008), which were reliably present across participants in both conditions (Kawabata Duncan et al., 2014; Woodhead et al., 2014; Fig. 1a). Six sources were identified and then modeled as equivalent current dipoles positioned symmetrically in each hemisphere, in a canonical brain (MNI) space, with prior mean location coordinates (x, y, z) at: OTC: −49, −62, −15 (left); 49, −62, −15 (right); anteromedial temporal lobe (AmTL): −35, −15, −30 (left); 35, −15, −30 (right); and IFC: −42, 30, −2 (left); 42, 30, −2 (right) (Fig. 1a; Campo et al., 2013). These sources were optimized at an individual level during DCM inversion using distributed dipoles, and the forward solution from the source localization (Moran et al., 2014). These anatomical regions are in fine agreement with the results from previous MEG studies of visual object naming (Clarke et al., 2011, 2013, 2014; Campo et al., 2013; Mousas et al., 2015). DCM analyses were performed using the most recent version of SPM12.
We followed the procedure described by Woodhead et al. (2014) as the structure of their experiment is very similar to ours. Accordingly, we first developed four competing model configurations that varied in number of sources and regional location. Configuration 1, included sources in left and right OTC only; Configuration 2, with left and right OTC, and AmTL sources; Configuration 3, with left and right OTC, and IFC sources; and Configuration 4, with left and right OTC, AmTL, and IFC sources (Fig. 1b). In all cases models were left hemisphere–right hemisphere symmetric (David et al., 2011), and the OTC bilaterally served as input region (Salmelin and Kujala, 2006; Heim et al., 2009; Schurz et al., 2014). All possible forward, backward, and homotopic lateral connections between sources were included in the models (Fig. 1b). These interconnections followed anatomical-functional evidences (for a similar approach, see Rudrauf et al., 2008). Specifically, occipitotemporal regions have been shown to be functionally connected with IFC during visual object processing (Bar et al., 2006), and anatomically by means of the inferior fronto-occipital fasciculus (Catani and Thiebaut de Schotten, 2008; Dick and Tremblay, 2012). AmTL has been shown to be functionally connected with occipitotemporal regions and with IFC (Pascual et al., 2015), underpinned by structural connections by means of the inferior longitudinal fascicle, and the uncinate fascicle, respectively (Catani and Thiebaut de Schotten, 2008; Duffau et al., 2013; Von Der Heide et al., 2013; Bouhali et al., 2014; Fan et al., 2014). Additionally, interhemispheric connections between homotopic regions have been extensively documented (Seacord et al., 1979; Clarke, 2003; Turken and Dronkers, 2011; Berlucchi, 2014). These models were inverted per participant for all trials (high- and low-demanding conditions) over the first 400 ms. We performed a random-effect (RFX) BMS procedure (Penny et al., 2004) to select the architecture that best explained the electromagnetic responses, based on the posterior exceedance probability (xp). This method quantifies how likely a specific model generated the data of a random subject in the context of a group of subjects, and is preferred when optimal models can vary across participants (Seghier et al., 2010; Stephan et al., 2010). Having established the model structure that provided the best fit to the data, in a second step, we estimated the modulatory effects on effective connectivity between sources for a different subset of connections (i.e., forward backward or both) induced by semantic demand (i.e., high vs low; Seghier et al., 2011; Woodhead et al., 2014). This was accomplished by Bayesian model averaging (BMA) over five successive 50 ms time windows, 1–200, 1–250, 1–300, 1–350, and 1–400 ms, covering the period where adding semantic-feature information improves modeling and classifying visual objects (Clarke et al., 2014). A requirement of the DCM analysis is that all time windows incorporate the time at which the stimulus was presented (Woodhead et al., 2014). This dynamic view of coupling has recently been coined as “chronnectome” (Calhoun et al., 2014).
Results
Behavioral results
Differences in performance accuracy between conditions were analyzed by a paired t test. Participants were more accurate on the low-demanding condition (M = 97.09, SD = 2.44) than in the high-demanding condition (M = 93.06, SD = 3.02; t(13) = 4.57, p < 0.001).
Source space analysis
We observed that both conditions consistently activated several regions in the ventral stream, but no differences were found between them (p < 0.005, uncorrected; Fig. 1a).
a, Axial views of the source localization for the grand-mean responses averaged over high- (left) and low-demanding conditions (right) projected into MNI voxel space and superimposed on the template structural MRI image. Sources of activity (middle), modeled as dipoles (estimated posterior moments and locations) superimposed on an MRI of a standard brain in MNI space, and their coordinates as included in the DCM analysis. b, Outline of the four DCM models for the first phase of effective connectivity analysis shown on axial brain schematics. The best model was selected using random-effects BMS (see Materials and Methods). c, Illustration of the models estimated and compared in the second phase of DCM analyses, which differed in the modulatory effects (i.e., forward, backward, and lateral) by the semantic demands. Modulated connections are shown with dashed arrows. Driving inputs are shown by stripped arrows. F, Forward; F-L, forward and lateral; B, backward; B-L, backward and lateral; FB, forward and backward; FB-L, forward, backward, and lateral.
BMS
BMS (RFX) revealed that the optimal architecture (i.e., the model with the highest xp) included sources in the left and right OTC, AmTL, and IFC sources (xp = 0.979; Fig. 1b).
Time-varying changes in effective connectivity between conditions
The winning architecture obtained in the previous step was used as the basic structure to estimate the modulatory effects of semantic demands on effective connectivity. As our main interest was to determine the role of recurrent dynamics, six competing models differed in the presence or absence of modulatory effects on backward connections. This allowed us to test whether the observed evoked responses were best explained by a model operating in a forward manner or whether inclusion of backwards connections improved the explanation of the data (Garrido et al., 2007; Furl et al., 2014). Additionally, modulation of lateral connections linking homologous areas was also included. Differences between conditions were evaluated by allowing connections to be modulated (David et al., 2011). Accordingly, six competing models, each representing a different way on how connections between sources are modulated, were fitted to each participant's ERF data (Auksztulewicz and Friston, 2015): forward; forward plus lateral; backward; backward plus lateral; forward and backward; forward and backward plus lateral (Fig. 1c). Finally, to explore the dynamics in the effective connectivity within the network as a function of task demands, this is, time-varying changes on functional architecture configuration and on the connectivity strengths (Calhoun et al., 2014; Clarke et al., 2014; Fedorenko and Thompson-Schill, 2014; Murray et al., 2014; Woodhead et al., 2014), models were specified and inverted separately from stimulus onset to a variable poststimulus time, ranging from 200 to 400 ms, in 50 ms steps (Garrido et al., 2007; Woodhead et al., 2014). BMS was done at the family level (Penny et al., 2010). In this case, the xp values represent the evidence of each family of models instead of the evidence of each individual model (Seghier et al., 2011). We defined three families that grouped models according to the dominant direction of modulatory effects with no overlap: forward (Fwd), backward (Bwd), and forward and backward (Fwd-Bwd; Garrido et al., 2007; Schurz et al., 2014; Fig. 2b). A BMA procedure (Penny et al., 2010) was then used to compute posterior means of connectivity parameters for each subject (Seghier et al., 2011). BMA estimate the average strength of the connectivity parameters of all the models tested, weighted by their posterior probabilities, so that the contribution of models with weak evidence is minimized (Seghier et al., 2011). This procedure generates a distribution of the model parameters that is proportional to the likelihood of each model given the data (Richardson et al., 2011). From the different sets of parameters estimated by DCM, we focused on modulatory parameters that measure the changes in effective connectivity induced by the experimental conditions (Cardin et al., 2011; Kawabata Duncan et al., 2014). Hence, parameters values for the modulations calculated during the BMA were obtained for each subject and extracted to make statistical inferences at the group level with repeated-measures ANOVA.
Random-effects BMS in five time windows. The bar graphs plot the exceedance probability at the (a) model level and (b) family level. F, Forward; F-L, forward and lateral; B, backward; B-L, backward and lateral; FB, forward and backward; FB-L, forward, backward, and lateral.
Family selection results
BMS (RFX) revealed that the family that provides the best explanation of the data varied across time. In the 1–200 ms time window, the family Fwd was clearly superior to all other families (xp = 0.931). In the following three time-windows (1–250, 1–300, and 1–350 ms) the family Fwd-Bwd was considered the winning family (xp = 0.782, xp = 0.439, xp = 0.875, respectively; Fig. 2b). However, in the time window 1–300 ms, despite the fact that family Fwd-Bwd had the greatest xp, it could not be considered as clearly superior to the other families. Note that the best three models belonged to each of the three families, representing 79% of xp (Fig. 2a). This indicated that it was not possible to clearly distinguish between families (David et al., 2011). Thus, in this case, the three families were considered as winning families, and were used to extract the modulatory connections calculated during a BMA procedure (Seghier et al., 2011; for a similar procedure, see David et al., 2011). In the 1–400 ms time-window, family Fwd had again the greatest evidence (xp = 0.779; Fig. 2b).
Modulatory effects of semantic demands: model parameters results
We conducted a series of repeated-measures ANOVA with modulatory parameters as the dependent measures, and condition (high-demanding vs low-demanding), hemisphere (left and right), and latency (200, 250, 300, 350, and 400 ms) as the within-subject factors. Note that for backward connections, latency only had three levels (250, 300, and 350 ms). Analyses were performed using SPSS 18.0. Effects were considered statistically significant when p < 0.05 after Bonferroni correction. Accordingly, SPSS Bonferroni adjusted p values are quoted (Rytsar et al., 2011).
The only modulatory effect that reached significance was observed in the backward connection from IFG to OTC (Fig. 3). Specifically, a main effect of Condition was observed (F(1,13) = 5.12, p < 0.05; η2 = 0.283). Further comparisons indicated that there was a significant increase in the negative coupling in the backward connection from IFC to OTC in the high- (Mean = −0.085, SEM = 0.029) as compared with the low-demanding condition (Mean = 0.044, SEM = 0.032; p < 0.05 corrected). Although the coupling in the backward connection from IFC to OTC was negative in the high- compared with low-demanding condition over time, this was especially marked in the 1–350 ms time window, when there was a significant increase in the negative coupling in the backward connection from left IFC to left OTC in the high- (Mean = −0.23, SEM = 0.087) compared with low-demanding condition (Mean = 0.068, SEM = 0.086; t(13) = 2.32, p < 0.05). Additionally, we observed a trend for a negative correlation between the backward connection from left IFC to left OTC and performance in the high-demanding condition in the 1–350 ms time window (r = −0.49, p = 0.075).
Schematic representation of the BMA results in the time windows where recurrent interactions became active: 1–250, 1–300, and 1–350 ms. Arrows represent the fixed effective connectivity that is present in the system regardless of the modulatory effect. Normal arrows represent positive (i.e., excitatory) connections; dashed arrows represent negative (i.e., inhibitory) connections. Values represent the modulatory effect of semantic demands on connection strength. Values for the high-demanding condition are represented by normal font and for the low-demanding condition by italics. Thick lines represent significant modulatory effect of semantic demands. Note that the significant modulation between IFC and OTC was observed over time. Driving inputs are shown by stripped arrows. LH, Left hemisphere; RH, right hemisphere.
Discussion
Using DCM for evoked magnetic fields (David et al., 2006) we were able to determine the extent to which semantic demands modulate the dynamics in the recurrent interactions during visual object naming. Semantic demands were operationalized by differences in significant psycholinguistic variables of the objects to be named (Graves et al., 2007). Interestingly, we observed a variation in the modulatory effects over time. Specifically, the DCM family with modulatory effects on forward connection only (family Fwd) had greater evidence than families including modulatory effects on backward connections (families Bwd and Fwd-Bwd) in the 1–200 ms time window (DiCarlo et al., 2012). In the following three consecutive time windows (1–250, −300, −350 ms), modulatory effects on recurrent connections became active, as family Fwd-Bwd appeared as the most likely, especially in the 1–350 ms time-window. Modulatory effects on backward connections were no longer relevant in the 1–400 ms time-window, when family Fwd had again the greater evidence (Fig. 2). In addition, we also observed that the contribution of interhemispheric connections appears to be relevant, as in each time window the configurations that better explain the data include this type of connection (Clarke, 2003). These interactions could contribute to coordinate and integrate specific activity generated by both hemispheres (Mima et al., 2001; Seghier et al., 2011); for instance a conceptual versus physical hemispheric specialization (Harel et al., 2014). The observed time course of feedforward and backward connections is largely consistent with the notion of an initial feedforward sweep (Thorpe et al., 1996; Riesenhuber and Poggio, 1999) enabling a coarse semantic processing (Wu et al., 2015), followed (>200 ms) by a recurrent processing supporting the formation of increasingly complex semantic representations (Lamme and Roelfsema, 2000; Hochstein and Ahissar, 2002; Schendan and Maher, 2009; Clarke et al., 2011, 2014; Ghodrati et al., 2014; Harel et al., 2014; Khaligh-Razavi and Kriegeskorte, 2014). Of particular interest to current findings, a recent MEG study of visual object recognition showed a similar timing of feedforward and backward interactions, with the latter being observed at 210–250 ms, but not earlier, when feedforward were predominant (Ahlfors et al., 2015). A similar modulation of the contribution of backward connections to evoked responses as a function of peristimulus time was demonstrated by Garrido et al. (2007) during a mismatch negativity task, when backward connections became essential after 220 ms. Additional support comes from a study that interfered visual recognition at 100 and 220 ms by applying transcranial magnetic stimulation over the occipital cortex (Camprodon et al., 2010). Domain or basic level naming was impaired at different moments, concluding that the 220 ms time window corresponded to a recurrent stage of visual processing (Roelfsema et al., 2002; Heinen et al., 2005; Wokke et al., 2012; Wu et al., 2015).
When focusing on specific interactions between sources, crucially, we observed that semantic demands significantly modulated the involvement of the positive (i.e., excitatory) recurrent processing from IFC to OTC over time (Fig. 3). Recurrent interactions involving IFC and posterior brain regions have been reported in several studies of visual object processing (Freedman et al., 2003; Bar et al., 2006; Ghuman et al., 2008; Schendan and Maher, 2009; Buffalo et al., 2010; Gilbert and Li, 2013; Harel et al., 2014; Murray et al., 2014). Top-down influences of IFC are considered to modulate the activity in posterior regions by biasing processing only to the most likely candidates, and thus facilitate detailed recognition of objects (Fenske et al., 2006; Schendan and Maher, 2009; Clarke et al., 2011; Trapp and Bar, 2015). Specifically, Bar et al. (2006; Chaumon et al., 2014) showed an enhanced phase synchronization between orbitofrontal cortex and the fusiform gyrus during an object recognition task. Differentially from this study, we used a method (DCM) that allowed us to examine the directionality of the causal interactions among brain regions, and to determine whether they are inhibitory or excitatory (Cardin et al., 2011). Our results showed that increased semantic demands caused a negative modulation in the excitatory backward connection from IFC to OTC over time, which was more evident in the left hemisphere in the 1–350 ms time window. Interestingly, we found that performance on the high-demanding condition showed a trend (r = −0.49, p = 0.075) to be negatively correlated with the backward connection from left IFC to left OTC in the 1–350 ms time window, such that more negative modulation was associated with better performance. Inhibitory effects in top-down signals from high level areas to lower level areas have been observed during visual processing of meaningful stimuli (Chen et al., 2009; Cardin et al., 2011) and lexical decision (Deng et al., 2012; Xu et al., 2015), that could result in more efficient processing (Ghuman et al., 2008). Recently, O'Reilly et al. (2013) described a biological model of the ventral visual pathway that provides evidence about the importance of recurrent inhibitory mechanisms in visual object recognition in situations that impoverished the image, that is, situations in which recognition is more difficult (Tang et al., 2014). Less frequent, atypical, unfamiliar and/or late-acquired concepts are thought to form weaker semantic associations and have less-detailed representations than more frequent/familiar or early acquired ones (Hirsch and Funnell, 1995; Lambon Ralph et al., 1998; Rogers et al., 2004; Woollams et al., 2008; Jefferies et al., 2009; Lambon Ralph et al., 2010; Hoffman et al., 2012; Woollams, 2012; Campo et al., 2014). Therefore, the role of the suppressive effect of the backward connection from prefrontal to posterior object-sensitive regions in the high-demanding condition could be to disambiguate the identity of the object to select one object concept from other exemplars (Walther and Koch, 2007; Gerlach and Marques, 2014), thus biasing processing only to the most likely candidates (Schendan and Kutas, 2002; Moss et al., 2005; Schendan and Maher, 2009; Clarke et al., 2011; Deng et al., 2012; Bruffaerts et al., 2013; Wright et al., 2015). This is, in the high-demanding condition object individuation (Gerlach and Marques, 2014) requires a greater inhibition of unintended exemplars (Deng et al., 2012). Compatible with this idea are the results from another study by Clarke et al. (2013), in which they suggested that the selection and retrieval of weakly correlated semantic information places greater demands on the conceptual system, driving an increased recruitment of frontal regions.
In summary, we used a method (i.e., DCM) that allows to understand how information propagates through brain regions (Kahan and Foltynie, 2013) by estimating the directionality of the causal interactions among brain regions and their associated connectivity strengths (Bianchi et al., 2013). Crucially, we explored the variation in the modulatory effects of semantic demands during visual-object naming across time, thus providing a dynamic measure of connectivity (Calhoun et al., 2014). We focused on interactions between brain regions across the ventral pathway (i.e., long-range interactions; Bressler and Richter, 2015), whereas recent studies have explored feedback influences as a function of behavioral demands among several occipital areas (Bastos et al., 2015). First, we observed a graded involvement of backward connections, which is largely consistent with the notion of an initial feedforward sweep enabling a coarse semantic processing, followed (>200 ms) by a recurrent processing supporting the formation of increasingly complex semantic representations (Hochstein and Ahissar, 2002; Schendan and Maher, 2009; Clarke et al., 2011, 2014; Harel et al., 2014; Khaligh-Razavi and Kriegeskorte, 2014). Second, we found that semantic demands caused a suppressive effect in the excitatory backward connection from IFC to OTC over time, which was remarkably evident in the 1–350 ms time window. Thus, current results complement those from previous studies underscoring the role of IFC as a common source of top-down modulation, which drives recurrent interactions with more posterior regions during visual-object recognition (Bar et al., 2006; Trapp and Bar, 2015). Crucially, our results revealed the inhibitory nature of this interaction in situations that place greater demands on the conceptual system (O'Reilly et al., 2013).
Footnotes
This work was supported by the Spanish Ministry of Science and Innovation, Grant SAF2011-27920 to I.G.-M., P.C. was supported by a Ramon y Cajal Fellowship from the Spanish Ministry of Science and Innovation (RYC-2010-05748), C.P. was supported by Spanish Ministry of Science and Education (AP2009-4131), and M.I.G. is supported by an Australian Research Council Discovery Early Career Researcher Award (DE1301011393) and the Australian Research council Centre of Excellence for integrative Brain Function (ARC Centre Grant CE140100007).
The authors declare no competing financial interests.
- Correspondence should be addressed to Dr. Pablo Campo, Departamento de Psicología Básica, Universidad Autónoma de Madrid, Campus de Cantoblanco, 28049 Madrid, Spain. pablo.campo{at}uam.es
References
In this issue
