Testing the effects of different top-down attention manipulations on local and global recurrent processing

Testing the effects of different top-down attention

manipulations on local and global recurrent processing

Several contemporary theories of consciousness postulate that one of the requirements of conscious perception is the integration of information through “recurrent processing”. This recurrent processing has been found to be affected by the attentional blink (AB), a manipulation known to affect top-down attention. The objective of the study described in this pre-registration is to see whether top-down attention manipulations, other than the AB, affect recurrent processing in the same way. This will be done with the use of the Kanizsa illusion and multivariate pattern analyses (MVPA), performed on recorded electroencephalogram (EEG) data of thirty participants performing multiple tasks that serve as top-down attention manipulations. These tasks are three feature detection tasks (one for every feature of the stimuli: collinearity, contrast, and illusion), a template matching task, a colour detection task, and a passive viewing task. A supplementary data analysis was done to replicate earlier findings and partially answer the main question of the pre-registered study. The findings of this analysis did not support earlier findings and expectations but due to statistical impairments (e.g. confounding factors and too little power); no conclusions regarding the main question could be made.

Keywords: access consciousness; phenomenal consciousness; recurrent processing; attentional blink;

Kanizsa illusion

everal contemporary theories of consciousness postulate that one of the requirements of conscious perception is the integration of information through “recurrent processing”. To clarify, when stimuli are translated into neural signals, this information is processed in the consecutive stages of the cortical hierarchy, which is called “feedforward processing” (Lamme and Roelfsema, 2000). Recurrent processing (via recurrent connections) can occur as soon as this information has reached the next stage, that is, information is sent back to hierarchically lower stages, or stages on the same level (Lamme, 2003). Crucially, only a portion of the recurrent connections is available for recurrent processing: those that connect neurons that are activated bottom-up (Roelfsema, 2006). Consequently, the degree of bottom-up stimulus strength determines the presence of recurrent processing (Dehaene, Changeux, Naccache, Sackur, & Sergent, 2006). This recurrent processing can take place locally, within sensory modules, or globally, when information is spread to a parieto-frontal network (Fahrenfort, van Leeuwen, Olivers, & Hogendoorn, 2017). Here it remains in a durable reverberating state and becomes usable for several cognitive processors, including those responsible for working memory, verbal report, or motor behaviour. This spreading is called the global ignition and progresses in a non-linear self-amplifying manner (Mashour, Roelfsema, Changeux, & Dehaene, 2020). An important condition that has to be met for processing streams to be globally ignited is that they exceed a certain activation threshold. This can be achieved with the use of top-down attention because the amount of top-down attention determines the depth of feedforward

s

and recurrent processing. Thus, a sufficient amount of top-down attention is needed for processing streams to cross the threshold and be globally ignited (Dehaene et al., 2006).

Some of these theories try to capture unconscious and conscious processing in a four-stage model, relying on the two aforementioned factors: 1) bottom-up stimulus strength (weak or strong) and 2) top-down attention (absent or present) (Dehaene et al., 2006). One of the model’s stages is that of subliminal processing, which happens when stimuli have too little bottom-up strength for recurrent processing to occur, even when sufficient top-down attention is directed towards it (Dehaene & Changeux, 2011). Alternatively, when stimuli do have enough bottom-up strength but lack sufficient top-down attention directed towards it, then, the recurrent processing will not be globally ignited and will, therefore, remain local. Following the model, processing under these circumstances is called preconscious processing. These scientists who term local recurrent processing that has yet to be globally ignited as “preconscious processing”, thus attribute “true” consciousness uniquely to global ignition (access). This conceptual form of “true” consciousness is referred to as access consciousness (Dehaene et al., 2006). On the other hand, other scientists have stated that the main ingredient for consciousness is recurrent processing, whether it is globally ignited or not (Lamme, 2006). This would imply that solely the presence of local recurrent processing could already be used as a definition for consciousness. This conceptual form of consciousness is called

phenomenal consciousness, which is proposed to hold the contents of conscious experience without requiring representations to be cognitively accessed (Block, 2011).

One way of studying recurrent processing is by studying how visual illusions such as the Kanizsa illusion (Fig. 1) are processed. When looking at this illusion, neurons with an illusionary line in their receptive field respond as if there is a real line (Kok, Bains, van Mourik, Norris, & de Lange, 2016). Consequently, an illusionary triangle surface is perceived. These neurons must thus receive information from other neurons to relay information about the illusionary triangle. A useful technique for studying how information is processed is by using multivariate pattern analysis (MVPA; informally

known as decoding). By applying MVPA to time-resolved signals such as those acquired from EEG recordings, the time-course with which neural representations are activated can be characterized. In order to do this, classifiers have to be trained to discriminate between experimental conditions (e.g. Kanizsa and control) and tested to evaluate decoding performance. Classifiers can be tested on the same time points as they were trained on (on-diagonal decoding) but can also be tested on other time points that were used for training (off-diagonal decoding). Plotting both in a so-called temporal generalization matrix thus also allows one to see how neural representation develop in time but also how they generalize across time (e.g. a neural activation pattern reoccurs later in time). Because of the latter, off-diagonal decoding provides a useful way to study recurrent processing (King & Dehaene, 2014).

With the use of the Kanizsa illusion and two kinds of manipulations known to affect conscious processing, Fahrenfort and colleagues (2017) studied the neural correlates of recurrent processing by decoding the illusion with the use of electroencephalographic (EEG) measures. These two

manipulations were: 1) masking, which is thought to impair local recurrent processing (Fahrenfort, Scholte, & Lamme, 2007), and the attentional blink (AB), which is known to disrupt global recurrent processing (Sergent, Baillet, & Dehaene, 2005). Participants were shown two black target figures (T1

Figure 1. Examples of the Kanizsa illusion and a non-illusionary variant. An illusionary triangle emerges when the black Pac-Man symbols align.

and T2) within a rapid serial visual presentation (RSVP) that also included red distractors in between the presentation of the targets. The two targets were used presented with a short lag between them to induce an AB. By using a long lag between the two targets in the noAB condition, it was ensured that participants would not experience an AB. Furthermore, T2 could be either strongly masked or unmasked in both conditions. Participants had to indicate whether the targets contained a surface region (Kanizsa) or not (control) at the end of each trial. In addition to this task, participants also executed a 1-back task of which the data was used to train the classifiers in the decoding analyses. This was done to ensure that task and response-related processes (such as those required for conscious decision-making) were not captured in the process of decoding. To asses decoding performance, the classifier was tested on T2 data.

The experimenters showed that, when decoding the illusion using data from the 1-back task as training data, decoding performance peaked at ~264 ms. Consequently, the experimenters interpreted this peak as reflecting recurrent processing, as this is required for surface perception in the Kanizsa illusion. More specifically, this peak reflected local recurrent processing following from the fact that task and response-related processes were excluded from the decoding analyses as training data was obtained from the 1-back task. Moreover, this local recurrent processing peak (local peak) did not differ between the short and long lag conditions but disappeared in both AB conditions when stimuli were masked. As a result, local recurrent processing seemed to be unaffected by the amount of top-down attention. Interestingly, the behavioural accuracy did drop because of the induced AB whereas masking resulted in chance level accuracy. An explanation for the absence of the local peak when stimuli were masked could have been that masking abolishes all processing of the stimulus, thus resulting in a floor effect. The experimenters tested this by decoding high- vs. low-contrast stimuli and collapsing over short and long lag. They found that both masked and unmasked stimuli showed high decoding performance for contrast peaking at ~80-90 ms. Consequently, this peak was interpreted as reflecting contrast detection that still seemed to occur despite masking.

Regarding Fahrenfort and colleagues’ (2017) finding that local recurrent processing seemed to be unaffected by the amount of top-down attention, a possible cause could be their choice for training their classifier on data from the 1-back task. This follows from the fact that during a 1-back task, one has to memorise the full configuration of the image, therefore attention has to be equally divided over the different aspects of the image. However, in the main task, attention was fully focused on perceiving the illusion. Therefore, their training data differed from their testing data regarding the amount of attention focused on perceiving the illusion. This may have prevented the decoding analysis from having significantly different performances at ~264 ms for the short lag and long lag conditions. Furthermore, when Fahrenfort and colleagues (2017) further investigated the neural processes underlying the decrease in behavioural accuracy during the AB, they ran a decoding analysis, this time, training the classifier on T1 data and testing on T2 data, thus including data involving conscious decision making about the presence of a Kanizsa. Remarkably, these results did show an effect of changing top-down attention on local recurrent processing, as it affected the activity in the visual cortex in a 200- to 300 ms time frame, which is associated with local recurrent processing (Koivisto, Salminen-Vaparanta, Grassini, Revonsuo, 2016).

In addition, they found that reducing top-down attention lowered the ability of the illusion to be decoded at another instance when decoding performance peaked as well, namely at ~406 ms. This is around the time frame of the P300 at 400-500 ms, an event-related potential (ERP) that has been frequently associated with cognitive processes requiring global recurrent processing such as attention and working memory. For example, Gebuis and Reynvoet (2013) found that the amplitude

of the P300 was decreased in passive viewing (i.e. passively looking at stimuli without any task demands) when compared with active processing of numerosity. As this effect could not be caused by response related processes, this peak seemed to reflect general cognitive processes that function differently when actively processing instead of passive viewing. Because of the temporal adjacency with the P300, and the fact that the classifier was trained on T1 data, thus no response related processes were involved, the ~406 ms peak was interpreted as reflecting global recurrent processing. Moreover, this global peak disappeared as well because of masking. Crucially, it seemed that the amount of top-down attention affected behavioural accuracy and the global recurrent processing peak (global peak) more than the local peak. This implies that in the total absence of top-down attention, local recurrent processing of the illusion would still take place, whilst global recurrent processing and the ability to distinguish between control and illusion would disappear.

Regarding the effects of the AB on local recurrent processing, there thus seems to be an inconsistency in the results acquired from decoding with training data from the 1-back task and from decoding with training data from the main task. In addition, it is yet unknown if top-down attention manipulations, other than the AB, would affect local and global recurrent processing in the same way as the AB did in the aforementioned study. To answer this question, this study will investigate how different top-down attention manipulations affect the decodability of the Kanizsa illusion.

This will be realised by decoding the data that will be acquired from measuring brain activity with EEG whilst participants execute six different tasks, involving the detection of certain features (Kanizsa illusion, collinearity, and contrast) present in the figures. These tasks differ among each other regarding the amount of top-down attention participants have for the features in the figures. In order to do this, features will be made relevant in one task and irrelevant in other task. By making contrast as well as collinearity task-relevant and task-irrelevant, it will be possible to investigate the influence of top-down attention manipulations on the decodability of these features as well. It has to be noted however, that the collinearity feature will not be covered in this project. These particular tasks are called 1 task-relevant feature (TRF) tasks and will have a version for every feature. Furthermore, there will be one task in which participants will have to divide their attention equally over the three features (3 TRF task) as well as one task in which participants will focus on none of the features at all (0 TRF Task). At last, there will be a passive viewing (PV) task in which participants will have no task demands. Following from the fact that the PV affected the P300, and that the P300 has been associated with global recurrent processing, the PV will be seen as an attention manipulation. Given the fact that the trials in these tasks will only have one target figure (as opposed to having T1 and T2), k-folding will be used during decoding to ensure that the training and testing data are similar

regarding the amount of attention focused on perceiving the features. Behavioural accuracy will be measured to evaluate the effectiveness of the top-down attention manipulations.

Additional data will be analysed to replicate the findings of Fahrenfort and colleagues (2017) and partially answer the main question of the study for which this pre-registration is aimed for. The analysis is provided as supplementary information at the end of this pre-registration.

Hypotheses

Behavioural manipulation checks

As stated earlier, to ensure the effectiveness of the top-down attention manipulation, behavioural accuracy will be measured. This will be used to assess whether tasks were of appropriate difficulty and participants were sufficiently motivated for the top-down attention manipulation to take effect.

M1: in tasks with an appropriate difficulty level, sufficiently motivated participants will perform

significantly higher than chance level.

EEG manipulation checks

To check whether the on-diagonal decoding performance peaks that are thought to reflect local and global recurrent processing really do reflect recurrent processing, off-diagonal decoding performance of the features in their respective 1 TRF task will be used as manipulation checks. This will be done by looking whether the off-diagonal decoding performance patterns resemble recurrent activation patterns (i.e. activation patterns that reoccur later in time).

M2.1: the off-diagonal decoding pattern in the temporal generalization matrix acquired from

decoding the illusion in its 1 TRF task will resemble a recurrent activation pattern.

M2.2: the off-diagonal decoding pattern in the temporal generalization matrix acquired from

decoding contrast in its 1 TRF task will not resemble a recurrent activation pattern.

M2.3: The off-diagonal decoding pattern in the temporal generalization matrix acquired from

decoding collinearity in its 1 TRF task will not resemble a recurrent activation pattern.

Top-down attention manipulations and illusion decodability

Based on the results of Fahrenfort and colleagues (2017), it is hypothesized that decreasing top-down attention for the illusion will proportionately reduce the local and global recurrent processing of the illusion. It is further hypothesised that the decrease in decodability will be larger for the global recurrent processing of the illusion than for the local recurrent processing of the illusion.

H1.1a: the on-diagonal decoding peaks, respectively reflecting local and global recurrent processing

of the illusion, will be highest during the illusion 1 TRF task, second highest during the 3 TRF task, third highest during the PV task, fourth highest during the other TRF tasks and lowest during the 0 TRF task.

H1.1b: the off-diagonal decoding patterns, respectively reflecting local and global recurrent

processing of the illusion, will be strongest during the illusion 1 TRF task, second strongest during the 3 TRF task, third strongest during the PV task, fourth strongest during the other TRF tasks and lowest during the 0 TRF task.

H1.2a: the on-diagonal decoding peaks, reflecting global recurrent processing of the illusion, will

decrease more than those representing local recurrent processing when comparing the illusion 1 TRF task with the 3 TRF task, the PV task, the other TRF tasks and the 0 TRF task.

H1.2b: the off-diagonal decoding patterns, reflecting global recurrent processing of the illusion, will

decrease more than those representing local recurrent processing when comparing the illusion 1 TRF task with the 3 TRF task, the PV task, the other TRF tasks and the 0 TRF task.

Top-down attention manipulations and contrast decodability

Following from the fact that the amount of top-down attention determines the depth of feedforward and recurrent processing, it is hypothesised that this will also increase the decodability of contrast.

H2: the on-diagonal decoding peak reflecting contrast detection will be greatest in height during the

contrast 1 TRF task, second greatest during the 3 TRF task, third greatest during the PV task, fourth greatest during the other TRF tasks and least greatest during the 0 TRF task.

Methods

Planned sample

Thirty subjects will participate. We will recruit them via the University of Amsterdam’s lab.uva.nl. People that have bad vision that is not corrected with glassed or lenses, have colour-blindness, or have a diagnosis or history of a psychiatric or neurological disorder are not allowed to participate. The allowable age range will be between 18 and 30 years.

Participants will be awarded with either research credits (one credit per hour) or money (€10 per hour) for their assistance. The behavioural laboratories at the University of Amsterdam will be used for testing. After 30 subjects, data collection will stop irrespective of eventual data exclusions during the data analysis.

Exclusion criteria

The data from tasks where participants failed to achieve a sufficient behavioural accuracy (i.e., significantly above chance level), will be excluded.

Figures

In this experiment, eight black target figures will be used and will be presented with Presentation (Neurobehavioral Systems). These figures will be presented on a white background on a 1920 x 1080 pixels monitor of 100 Hz refresh rate. Every figure will have three features, that each have two levels. These features and their levels will be collinearity (absent / present), illusion (absent / present), and contrast (high / high-low). With low-high and low-high-low is meant the orientation of the Pac-Man symbols, as can be seen in Fig.2. Each of the figures will have a unique combination of these feature levels in order to have eight distinctive figures.

Procedures

Each participant will run through all the six tasks, three of which

task. The order in which participants run through tasks will be counterbalanced across participants. In every task, the figures will be presented within a RSVP for 40 ms each. Participants will have a limited amount of time after figure onset to respond accordingly, following the task demands, by pressing the correct button with their left or right index fingers (one button per finger). This time duration will be randomised between 900, 950, 1000, 1050, and 1100 ms to ensure that participants will not be able to prepare their button press. To prevent the response-related processes from contaminating the decoding analyses, the meaning of the buttons will be counterbalances between blocks. For every task, participants will be required to finish 600-700 trials that will be divided over eight blocks, with all the eight figures being presented 10 times in a block.

1 TRF tasks

In the 1 TRF tasks, participants will be instructed to indicate whether the figure contained level 1 or 2 of the task-relevant feature. Thus, in the illusion TRF task, participants will have to indicate in every trial whether the figure contained an illusion or not. The same goes for the collinearity TRF task, participants will indicate whether the figure contained collinearity or not. For the contrast TRF task, however, participants will have to indicate if the figure had a low-high or a high-low configuration.

3 TRF task

For the 3 TRF task, participants will perform a template matching (TM) task. In this task, participants first see a template figure that they have to memorise. In the subsequent trials, they will have to report whether the figure shown matches the template figure or not. Each template figure will have its own block. The match-mismatch ratio will be 1:1 as every figure is presented equally often in the tasks.

0 TRF task

A colour discrimination (CD) task will be used for the 0 TRF task. During this task, participants will be asked to focus on a fixation point that changes colour after a random amount of time, and report whether the colour has changed to red or green by pressing the correct button. The amount of time after which the colour will change, randomises between 1500 and 6000 ms, with the changing in one frame. Subsequently, participants will have 1500 ms to respond. The choice for randomising the amount of time after which the colour will change is because this will make the colours totally unrelated to the figures. This will ensure that the figures and their features really are task irrelevant.

PV task

In the PV task, the only task demand for the participants is that they look at the figures. They will not need to respond, nor will they have to focus on specific features.

Analysis plan

Variables

The independent variables and their levels that this experiment will have are:

1. Task and with that the amount of attention for a feature (illusion TRF task: full attention / collinearity TRF task: full attention / contrast TRF task: full attention / TM task: divided attention / PV task: no task demand / CD task: no attention)

3. Illusion (absent / present) 4. Contrast (low-high / high-low) 5. Peak (local peak / global peak)

These variables are within participants and are in an orthogonal relationship with each other. The dependent variable used for EEG classification accuracy will be the area under the curve (AUC) and for the behavioural accuracy the hit rate (HR) – false alarm (FAR) rate. A more detailed explanation about the AUC and the HR – FAR will be given in the sections covering the EEG and the behavioural analyses.

Behavioural analysis

Responses will be labelled as hits (feature correct), misses (feature incorrect), correct rejections (control correct), and false alarms (control incorrect).

As earlier stated, the HR - FAR refers to the hit rate – the false alarm rate and will serve as a parameter for participant’s behavioural accuracy.

EEG data collection and pre-processing

EEG data will be acquired at 1024 Hz and recorded from 64 scalp electrodes placed according to the 10/20 system and two external electrodes attached to each earlobe, using ActiveTwo system (BioSemi, Amsterdam, the Netherlands). Data will be re-referenced to the average activity recorded at the earlobes, downsampled to 512 Hz, and epoched between -500ms before, and 1000 ms after stimulus onset. Furthermore, a baseline correction will be used.

With the use of the ft_artifact_zvalue muscle artefact detection function from the FieldTrip toolbox (Fries & Schoffelen, 2011), trials that had muscle artifacts could be excluded.

EEG MVPA

The decoding analysis will be performed with the use of Matlab (MathWorks) and the ADAM toolbox (Fahrenfort, van Driel, van Gaal, & Olivers, 2018). A backward decoding classification algorithm, using a 10-fold cross-validation scheme will be used for every participant’s data. First, the order in of trials will be randomized to remove information about the order in which trials were acquired during the experiment. Next, the raw EEG data will be divided into 10 equally sized subsets. Afterwards, to differentiate between the two levels of the features, a linear discriminant classifier will be trained using a 1/10 part the data and tested on the remaining 9/10 of the data to ensure independence of training and testing sets. This process of splitting data into subsets will be repeated k times until all the data is used for testing once whilst using different data in every train-test cycle to avoid circularity. By executing the cross-validation for each time sample of the EEG data, the evolution of classification accuracy over time will be acquired. To determine whether classification accuracies differ significantly from chance across subjects, one-sample two-sided t-tests will be used. For these intervals of significant decoding, the used t-tests will be corrected for multiple comparisons with the use of group-wise cluster-based permutation testing (1000 iterations, at a threshold of 0.05). In this procedure, the sum of t-values in the observed cluster of temporally contiguous significant data points is compared to the sum of t-values in a cluster of temporally contiguous significant data points obtained under random permutation. The p-value, with which the significance of a cluster in the

observed data will be evaluated, stands for the number of times that the sum of t-values of the cluster under permutation exceeds that of the observed cluster, divided by the number of iterations. Relevant time windows will be selected by inspecting the decoding accuracies acquired from K-folding T1 data. The maximum AUCs value in these time windows will be used.

Statistical analyses

The data analysis will be performed using R studio (R Core Team), with a car package (Fox & Weisberg, 2011). As the main tests to be used will be repeated-measures analyses of variance (rANOVAs), the assumption of spehricity will be tested. To compare tasks regarding their effects on the decodability of illusion and contrast in the post hoc, paired t-tests will be used. The alpha values of these tests will be corrected with the Bonferroni correction.

Manipulation checks

M1: To check whether the tasks are of appropriate difficulty and participants are sufficiently

motivated, participants HR - FAR will be tested against chance level with the use of one-sample t-tests.

M2.1, M2.2, and M2.3: determining whether the temporal generalization matrices acquired during

the off-diagonal decoding of the features in their 1 TRF task resemble significant recurrent activation patterns, will be done by visual inspection.

Top-down attention manipulations and the decodability of illusion and contrast

H1.1: To see if decreasing top-down attention for the illusion will proportionately reduce its local and

global recurrent processing, main effects of the tasks on the on- and off-diagonal illusion decoding AUCs will be tested. To find the main effects, two (one for local recurrent processing and one for global recurrent processing) repeated-measures analyses of variance (rANOVAs) will be used with factor task (illusion TRF vs collinearity TRF vs contrast TRF vs TM vs PV vs CD). These two rANOVAs will be performed for both the on-diagonal decoding AUCs and the off-diagonal decoding AUCs.

H1.2: To see whether the top-down attention manipulations affect global recurrent processing more

than local recurrent processing, interaction effects between the tasks and the peaks on the on- and off-diagonal illusion decoding AUCs will be tested. This will be done with two (one for on-diagonal decoding and one for off-diagonal decoding) 6 x 2 rANOVAs with the factors task (illusion TRF vs collinearity TRF vs contrast TRF vs TM vs PV vs CD) and peak (local peak vs global peak).

H2: To test whether decreasing top-down attention for contrast will proportionately reduce contrast

detection, main effects of the tasks on on-diagonal contrast decoding AUCs will be tested with a rANOVA with factor task (illusion TRF vs collinearity TRF vs contrast TRF vs TM vs PV vs CD)

References

Block, N. (2011). Perceptual consciousness overflows cognitive access. Trends in cognitive

sciences, 15(12), 567-575.

Dehaene, S., & Changeux, J. P. (2011). Experimental and theoretical approaches to conscious processing. Neuron, 70(2), 200-227.

Dehaene, S., Changeux, J. P., Naccache, L., Sackur, J., & Sergent, C. (2006). Conscious, preconscious, and subliminal processing: a testable taxonomy. Trends in cognitive sciences, 10(5), 204-211.

Fahrenfort, J. J., Van Driel, J., Van Gaal, S., & Olivers, C. N. (2018). From ERPs to MVPA using the Amsterdam decoding and modeling toolbox (ADAM). Frontiers in neuroscience, 12, 368.

Fahrenfort, J. J., Van Leeuwen, J., Olivers, C. N., & Hogendoorn, H. (2017). Perceptual integration without conscious access. Proceedings of the National Academy of Sciences, 114(14), 3744-3749.

Fahrenfort, J. J., Scholte, H. S., & Lamme, V. A. (2007). Masking disrupts reentrant processing in human visual cortex. Journal of cognitive neuroscience, 19(9), 1488-1497.

Gebuis, T., & Reynvoet, B. (2013). The neural mechanisms underlying passive and active processing of numerosity. NeuroImage, 70, 301-307.

King, J. R., & Dehaene, S. (2014). Characterizing the dynamics of mental representations: the temporal generalization method. Trends in cognitive sciences, 18(4), 203-210.

Koivisto, M., Salminen Vaparanta, N., Grassini, S., & Revonsuo, A. (2016). ‐ Subjective visual awareness emerges prior to P3. European Journal of Neuroscience, 43(12), 1601-1611. Kok, P., Bains, L. J., van Mourik, T., Norris, D. G., & de Lange, F. P. (2016). Selective activation of the

deep layers of the human primary visual cortex by top-down feedback. Current Biology, 26(3), 371-376.

Lamme, V. A. (2003). Why visual attention and awareness are different. Trends in cognitive

sciences, 7(1), 12-18.

Lamme, V. A. (2006). Towards a true neural stance on consciousness. Trends in cognitive

sciences, 10(11), 494-501.

Lamme, V. A., & Roelfsema, P. R. (2000). The distinct modes of vision offered by feedforward and recurrent processing. Trends in neurosciences, 23(11), 571-579.

Mashour, G. A., Roelfsema, P., Changeux, J. P., & Dehaene, S. (2020). Conscious Processing and the Global Neuronal Workspace Hypothesis. Neuron, 105(5), 776-798.

Oostenveld, R., Fries, P., Maris, E., & Schoffelen, J. M. (2011). FieldTrip: open source software for advanced analysis of MEG, EEG, and invasive electrophysiological data. Computational

intelligence and neuroscience, 2011.

Roelfsema, P. R. (2006). Cortical algorithms for perceptual grouping. Annu. Rev. Neurosci., 29, 203-227.

Sergent, C., Baillet, S., & Dehaene, S. (2005). Timing of the brain events underlying access to consciousness during the attentional blink. Nature neuroscience, 8(10), 1391-1400.

Supplementary data analysis

As stated in the pre-registration, this supplementary data analysis was done to replicate the findings by Fahrenfort and colleagues (2017). The most important finding that was tried to be replicated was that both local recurrent processing and global recurrent processing were affected by the amount of top-down attention with global recurrent processing being affected more than local recurrent processing. Furthermore, the question whether top-down attention manipulations, other than the AB, would affect local and global recurrent processing in a similar way as the AB was tried to be partially answered by doing this data analysis. The data that was analysed to supplement the pre-registration originates from a running study that is similar to the study of Fahrenfort and colleagues (2017) with a few exceptions. Most notably, this study’s main aim is to investigate the effects of Memantine, a drug hypothesised to affect recurrent processing by blocking the NMDA receptor, on conscious processing. Similarly, to Fahrenfort and colleagues (2017), this is done by studying its effects on the decodability of the Kanizsa illusion whilst implementing masking and the AB. However, to train the participants, the TM task is used in the Memantine study in contrast to the 1-back task used in Fahrenfort and colleagues’ (2017) study.

The decoding analysis used on the data from the Memantine study was performed in the same way as described in the pre-registration. This data-analysis used the AUC and the proportion of correct trials (CTP) as parameters for respectively EEG decoding accuracy and behavioural accuracy as opposed to the HR – FAR, used in as the study of Fahrenfort and colleagues (2017). In contrast to the pre-registration, all the AUCs that were used this analysis originate from on-diagonal decoding, as opposed to some of the AUCs described in the pre-registration that originate from off-diagonal decoding.

Data analyses

To see whether masking leaves the contrast detection peak unaffected in the Memantine study, a rANOVA was used with factor masking (unmasked vs. masked) to test for significant effects of masking on the contrast AUCs whilst collapsing over the lags. To replicate Fahrenfort and colleagues’ (2017) findings of the effects of masking and the AB on local and global recurrent processing and the accuracy with which it is being discriminated from control, main effects of, and interaction effects between masking and the AB on the illusion AUCs and the CTPs were tested. This was done by using a rANOVA with the factors masking (unmasked vs. masked) and AB (short vs. long lag) for both the local and global peak. To investigate whether global recurrent processing was affected more by the AB than local recurrent processing in the Memantine study, interaction effects between the AB and the peak on the illusion AUCs were tested. With the use of a rANOVA with the factors AB (short vs. long lag) and peak (local vs. global peak), this was realised. By using a rANOVA with the factors AB (short vs. long lag) and measures (local peak vs. behavioural accuracy), interaction effects between the AB and the peaks on the illusion AUCs and the CTPs were tested as well to see whether behavioural accuracy was affected differently than the local recurrent processing.

With the use of the data from the TM task in the Memantine study, the question whether the top-down attention manipulations described in the pre-registration affect local and global recurrent processing similarly as the AB, could be answered partly. This was done by comparing the AUCs of the

local and global peaks separately between the unmasked-long lag condition of the illusion detection task, the TM task, and the unmasked-short lag condition of the illusion detection task. As there is much attention in the long lag condition, divided attention in the TM task, and little attention in the short lag condition for the illusion, the illusion AUCs were expected to decrease in the same order, with the global illusion AUCs decreasing more than the local illusion AUCs, following the hypotheses in the pre-registration. To test for main effects of the attention manipulation, a rANOVA was used with factor manipulation (unmasked-long lag AB vs. TM task vs. unmasked-short lag AB) for each of peaks. In order to test for interaction effects between the attention manipulations and the peaks on the illusion AUCs, a rANOVA was used as well with the factors manipulation (unmasked-long lag AB vs. TM task vs. unmasked-short lag AB) and peak (local vs. global).

Results

T1 K-folding

When decoding the illusion using K-folded T1 data, Kanizsa vs. control decoding accuracy peaked at ~175 ms as well as at ~321 ms and low-high vs. high-low decoding accuracy peaked at ~89 ms. Using the same rational as described in the pre-registration these peaks were interpreted as reflecting respectively local recurrent processing, global recurrent processing, and contrast detection (Fig. 3).

T1 training – T2 testing: AB and masking

No significant main effects were found for masking on the contrast AUC [F(1,5) = 1.658;

P = 0.239] as can be seen in Fig.4.

In contrast, a significant main effect of masking [F(1,5) = 321.134; P < 10-5_{], but no} main effect of AB [F(1,5) = 2.507; P = 0.174] or interaction effect between AB masking [F(1,5) = 0.949; P = 0.375] on the local illusion AUC were found (Fig. 6). These results were confirmed in the post hoc, with the local illusion AUC being significantly higher in the unmasked conditions than in the masked conditions for both the long lag [t(5)=3.140; P < 0.05] and the short lag [t(5)=3.596; P < 0.05].

Likewise, a significant main effect of masking [F(1,5) = 306.574; P < 10-4_{], but no} main effect of AB [F(1,5) = 3.199; P = 0.134]

Figure 3.EEG decoding accuracies over time for the illusion (Left) and contrast (Right) when using K-folded T1 for the classifier. Light coloured line and the light coloured area around it represent mean ± SEM. Darker lines represent significant decoding (P <0.05, cluster-based permutation test).

Figure 4. Graphical representation of the EEG decoding accuracies at 89 ms in the unmasked and masked conditions when decoding contrast using T1 data to train, and T2 data to test the classifier. Individual data are presented in light colours , the mean as a darker line with error bars that represent mean ± SEM. ns, not significant (P > 0.05).

nor interaction effect between the AB and masking [F(1,5) = 5.259; P = 0.070] on the global illusion AUC were found. Following from the post hoc, the glob illusion AUC was found to be significantly higher in the unmasked conditions than in the masked conditions for both the long lag [t(5) = 7.170;

P < 0.001] and the short lag [t(5 )= 3.164; P < 0.05], but this effect was stronger for the long lag

(Fig.6).

As for the CTP however, a significant main effect of the AB was found [F(1,5) = 30.155; P < 0.01] in addition to a significant main effect of masking [F(1,5) = 975.551; P < 10-6_{]. A significant} interaction effect was found as well [F(1,5) = 39.647; P < 0.01] with the CTP being significantly higher in the long lag trials than in the short lag trials but only for unmasked trials [t(5) = 4.632; P < 0.01]. Similarly to the local and global illusion AUCs, CTP was found to be higher in the unmasked conditions than in the masked conditions for both the long lag [t(5) = 3.140; P < 0.05] and the short lag [t(5) = 3.596; P < 0.05] (Fig 5). Surprisingly, when investigating whether the AB affected the global peak and the behavioural accuracy more than the local peak, no significant interaction effects were found between the AB and peak factors [F(1,37) = 0.016;

P = 0.899] or between the AB and measure factors

[F(1,38) = 1.466; P = 0.233] on the illusion AUCs and CTP.

T1 training – TM and T2 testing: AB and TM

A

B

Figure 6. EEG decoding accuracies in the four experimental conditions when decoding the illusion using T1 data to train, and T2 data to test the classifier. (A) Graphical representation of the EEG decoding accuracies at 175 ms (Top) and at 321 ms (Bottom). Individual data are presented in light colours, the mean as a darker line with error bars that represent mean ± SEM. ns, not significant (P > 0.05), * P < 0.05, *** P < 0.001. (B) EEG decoding accuracies of the illusion over time. Light coloured line and the light coloured area around it represent mean ± SEM. Darker lines represent significant decoding (P <0.05, cluster-based permutation test).

Figure 5. Graphical representation of the behavioural accuracies at T2 in four experimental conditions. Individual data are presented in light colours, the mean as a darker line with error bars that represent mean ± SEM. ns, not significant (P > 0.05), ** P < 0.01, *** P < 0.001.

Figure 7. Graphical representation of the EEG decoding accuracies at 175 ms (Left) and at 321 ms (Right) in the unmasked long lag condition, TM task, and unmasked short lag condition when decoding the illusion using T1 data to train, and TM data and T2 data to test the classifier. Individual data are presented in light colours, the mean as a darker line with error bars that represent mean ± SEM. ns, not significant (P > 0.05), * P < 0.05.

For the local illusion AUC, no significant main effect was found for the manipulation factor [F(2,10) = 3.930; P = 0.055], whereas for the global illusion AUC, there was a significant main effect of the manipulations [F(2,10) = 8.084; P < 0.01]. In the post hoc findings (Fig. 7) however, as the global illusion AUCs did not differ significantly between the long lag condition and the TM task [t(5) = 2.360;

P = 0.070], the long lag condition and the short lag condition [t(5) = 1.348; P = 0.120], or the short lag

condition and the TM task [t(5) = 0.967; P = 0.190]. When investigating whether the manipulations affected the global peak more than the local peak, a significant interaction effect was found between the manipulation and peak factors [F(1,37) = 4.345; P < 0.05].

Discussion

Following from the results, we show that the replication of Fahrenfort and colleagues’ (2017) decoding of the Kanizsa illusion and contrast was done successfully with the use of data from the Memantine study, resulting in decoding performance peaks that reflect local recurrent processing, global recurrent processing, and contrast detection. As no significant main effect of the AB on the AUCs of the local and the global peaks was, no supporting evidence for the findings of Fahrenfort and colleagues (2017) that the amount of top-down attention affects both local and global recurrent processing, with the latter in a higher degree, could be found. As in Fahrenfort and colleagues’ (2017) study, the potential explanation that masking might nullify all processing was not applicable to the finding that the local and global peak disappeared when stimuli were masked. This follows from the fact that the contrast detection peak in this decoding analysis was unaffected by masking, implying that some processing (i.e. that of contrast) still occurs when stimuli are masked. Remarkably, by using the TM task as an additional top-down attention manipulation, a significant effect of the

manipulations on the AUC of the global peak was found. However, this was not the case for the AUC of the local peak, nor was it the case for either of the peaks when only the effect of the AB was inspected. Therefore, it cannot be concluded that the TM as a top-down attention manipulation has similar effects on recurrent processing as the AB.

Notwithstanding the conclusions made, it is more likely that these results are caused by impairments of the data that was used, rather than from flaws in the theory behind the role of attention in recurrent processing and consciousness. One of these impairments might have been that half of the participants was given Memantine, a drug hypothesised to block the NMDA receptor thereby decreasing recurrent processing. Logically, this would decrease the accuracy with which a decoding analysis discriminates between Kanizsa and control using their neural representations, as the mechanism that differs between the processing of the two, namely recurrent processing, is impaired. Another important limitation of the used data was its sample size (n = 6). Because of this, the analyses lacked too much power to find significant results at places where it would have been expected to do so. This becomes clear from findings such as the one where, even though a main effect was only found for the AB on the CTP and not on the local illusion AUC, no significant

interaction effect between the factors measure and AB on the CTP and local illusion AUC was found. Likewise, the finding that no significant interaction effect was found between factors masking and AB on the global illusion AUC even though the global illusion AUC differed more between the unmasked and masked conditions for long lag trials than for short lag trials.