Rafał Szymanek
University of Amsterdam
MSc Brain and Cognitive Sciences, cognitive neuroscience track
When is visual awareness continuous and
when is it all-or-none?
ABSTRACT
In an influential paper, Sergent and Dehane (2004) claimed that visual awareness is all-or-none. This finding has recently been confirmed by Asplund and colleagues (2014) who showed that responses in a colour-detection task can be modelled as a mixture of guesses and correct trials of fixed precision. However, several theoretical approaches (Kanwisher, 2001), as well as empirical data (Ramsøy & Overgaard, 2004; Christensen, Ramsøy, Lund, Madsen, & Rowe, 2006; Sandberg, Bibby, Timmermans, Cleeremans, & Overgaard, 2011) suggest that awareness is a continuum. Here, we attempt to reconcile these differences by proposing that visual awareness depends on two processes. The bottom-up process depends on stimulus quality and has continuous characteristics, while the top-down process is contingent on attention and is dichotomous. We propose that differences in findings stem from using different paradigms: experiments employing the attentional blink are actually studying the top-down process, whereas research on backward masking is tapping into the bottom-up process. Therefore, findings from either of the paradigms cannot be generalized onto awareness as a whole. In the present study, we embedded the methodology of Asplund et al. (2014) in a backward masking paradigm. Unlike the original authors, who studied the attentional blink, we find that our findings cannot be explained by mere increase in the guess rate and that response precision is affected. The experiment provides support for our two-process model.
January – June 2017 Student ID: 11119438 ECTS points: 36
Supervisor: Prof. John-Dylan Haynes, Bernstein Center for Computational Neuroscience, Berlin Co-assessor: Dr. Simon van Gaal, University of Amsterdam
Introduction
Continuous or all-or-none?
It is a well-established finding in cognitive sciences that subliminal presentation can lead to changes in behavioural responses (Tulving & Schacter, 1990). In other words, presenting a stimulus in a way that prevents it from being consciously reported can nevertheless, under specific circumstances, affect the performance or reaction times in a cognitive task. Participants that have no explicit, reportable information about a stimulus can use implicit information to perform on a higher-than-chance level in subsequent tasks. This phenomenon has been observed in a multitude of tasks, ranging from basic identification and naming of primed stimuli (Tulving & Schacter, 1990) to more complex tasks like performing arithmetic (Sklar et al., 2012). Classically, a distinction has been made between unconscious (or subliminal) and conscious processing, with the latter being characterized by explicit, reportable knowledge.
More recently, a debate has emerged about whether there exist in-between states of graded awareness. In other words, do representations of visual stimuli enter consciousness suddenly, in an all-or-none, dichotomous fashion, or do they build up in a gradual, continuous manner. Several imaging studies have shown that an increase in subjective awareness of a stimulus leads to gradual increases in neural activity (Bar, Tootell, Schacter, Greve, & Fischl, 2001; Grill-Spector, Kourtzi, & Kanwisher, 2001). Furthermore, signal detection theory and connectionist models generally predict that awareness is continuous (Kanwisher, 2001).
However, some researchers have advocated for the dichotomous account. Sergent and Dehaene (2004) showed that people perceive visual stimuli in an all-or-none manner. They based their reasoning on the results of their experiments which used the attentional blink paradigm (AB). AB occurs when a series of visual stimuli are presented rapidly in a sequential manner, and the participant is asked to detect two targets occurring in the stream. Correct identification of the first target leads to a period lasting between 200 and 500 ms, during which presenting the second target does not cause a conscious identification (Broadbent & Broadbent, 1987). If the second target is presented after that period, it will be identified without trouble. AB is a standard paradigm, commonly used in consciousness research as a method of invoking loss of awareness. Sergent and Dehaene (2004) investigated whether the AB causes a complete, all-or-none loss of conscious access, or whether it degrades the information in a gradual manner. To answer that question, they employed a continuous visibility scale, asking participants to indicate on a vertical line the visibility of the second target, ranging from “not seen” to “maximal visibility”. They then varied the interval between the two targets, using 8 different intervals between 86 and 688 ms, with a step of 86 ms, and analysed the distribution of the responses on the visibility scale. They found that participants were primarily making use of the two ends of the visibility spectrum: either they had not seen the stimulus and the visibility was near 0%, or they had seen it clearly, and judged the visibility as close to 100%. This led the researchers to conclude that visual awareness works in an all-or-none fashion.
Similar conclusions were also reached by Del Cul and colleagues (2006). They showed a number, which served as a prime, in one of four locations on the screen. Then, after an interval ranging between 0-150 ms, they showed a target number, along with a mask composed of three letters in the location surrounding the location of the original prime. They then asked participants to perform a number comparison task on either the prime or the target, as well as to indicate the degree of
visibility on a continuous scale similar to the one used by Sergent and Dehaene (2004). Like in the study described above, they found that participants were using the visibility scale in an all-or-none fashion. What is worth noting, is that this particular version of backward masking resembles the AB. Because both the prime and target could be probed, participants had to pay attention to both. In effect, the prime and target could function as the first and second target of the AB, respectively, with detection of the prime causing a loss of access to the target. This paradigm might lead to different results than the AB, for example in terms of the threshold interval, due to the lack of a stimulus stream and the presence of a mask. At the same time, however, it is not a pure masking experiment. Other researchers have challenged these claims. Ramsøy and Overgaard (2004) criticised the lack of consideration of graded awareness in previous research and constructed a perceptual awareness scale (PAS), designed to give insight into the subjective quality of visual representations. Using this scale, they conducted a backward masking experiment. This is another commonly used technique in consciousness research. The presence of the mask overrides the content of awareness, rendering the target unconscious (Breitmeyer & Ogmen, 2000). In this experiment, participants saw simple shapes (triangles, circles and squares) with one of three basic colours (red, blue and green) displayed on the screen for between 16 and 192 ms, followed by a mask composed of all stimulus features. Participants were asked to report all the features of the target, as well as to indicate the quality of their experience using the 4-point PAS. The researchers found that the representation quality (or, the clearness of the perceptual experience), indicated by the PAS, was positively correlated with stimulus duration—participants were reporting stronger and stronger subjective conscious experience along with the increase in the objective stimulus strength. This led the authors to conclude that visual awareness is graded.
Further research conducted by the same team (Overgaard, Rote, Mouridsen, & Ramsøy, 2006) have further corroborated the continuous awareness claim. In their paper, Overgaard and colleagues attempted to dismiss the conclusions reached by Sergent and Dehaene (2004)—that visual awareness is all-or-none—using a PAS along with a backward masking experiment. They used textured displays composed of 91 oriented elements. Within the texture, a target consisting of four orthogonal elements was displayed in one of four possible locations. The stimulus was displayed for 12-96 ms, and was followed by a mask which was a uniform textured display, without the target. Figure 1 shows a sample stimulus used in the study. Using a regression model, the authors found that participants were using the scale in a linear, continuous manner. They conclude that this is evidence for a graded character of visual awareness.
Therefore, we can identify a discrepancy in the literature, with two camps arriving at opposite conclusions. Some researchers have tried to tackle this discrepancy by arguing that the results obtained by Sergent and Dehaene stemmed from the confidence scale that they used—namely, that a continuous 21-point scale with labels only on the extreme ends of the scale induce all-or-none responses (Nieuwenhuis & de Kleijn, 2011). However, this claim was tested by Pretorius, Tredoux and Malcolm-Smith (2016), who compared four variations of the same masking task, using a 3-, 4-, 7- and 21-point confidence scale. They transformed the responses from each scale to a 3-point scale corresponding to low awareness (less than 75% accuracy), intermediate awareness (75-97.5% accuracy) and high awareness (over 97.5% accuracy). They found that, while longer scales decreased the prevalence of graded (intermediate) awareness, they did not preclude its presence. In fact, using the same analysis as Sergent and Dehaene on the 21-scale version of their task, they found continuous responses. Therefore, the choice of the visibility scale cannot explain the differences in findings and an alternate explanation must be sought.
In this paper, I will argue that these differences come about because of another methodological choice, namely the use of either the AB, or AB-like paradigms on one side, and backward masking on the other. While most experiments supporting the graded account have used backward masking (Ramsøy & Overgaard, 2004; Overgaard et al., 2006; Sandberg, Timmermans, Overgaard, & Cleeremans, 2010; Sandberg et al., 2011), researchers advocating the all-or-none account have either relied on the AB paradigm (Sergent & Dehaene, 2004; Nieuwenhuis & de Kleijn, 2011, experiment 1) or a version of masking which resembles AB (Del Cul et al., 2006, 2007). This suggests that separate mechanisms might be employed in constructing visual awareness, with a dichotomous process responsible for the results observed in the AB experiments, and a continuous process responsible for the effects of backward masking.
The two-process model
I would like to propose a two-process model of visual awareness. According to this model, two processes are involved in constructing a conscious representation.
The first process can be described as a top-down mechanism that is guided by attention. It operates in a dichotomous fashion: only stimuli that have received a level of attention that surpasses a certain threshold are processed strongly enough to reach consciousness. This is the mechanism that is being manipulated in the attentional blink. Once the first item captures the participant’s attention, an interval of sufficient length is required so that enough attentional resources can be allocated to the second item.
The second process can be described as bottom-up, driven by stimulus strength or quality. Stimulus quality can be manipulated in several ways, all of which have been shown to induce the loss of conscious perception of the stimulus. Examples include decreasing the stimulus presentation, obstructing parts of the image, or superseding the image with a mask. This is a continuous process, wherein an improvement in the stimulus strength leads to an increase in subjective visibility, which can be equated with a more reliable conscious representation or a higher degree of stimulus awareness. The backward masking experiments described here, according to this model, tap into this process. Increasing the interval between the target and the mask, or increasing the on-screen duration of the target, thus leads to more sensory information about the target being available for further processing, which results in an increasing degree of visual awareness.
This model ties very well with several highly influential accounts of conscious processing, which describe the relation between attention and stimulus strength, and their implications on consciousness, namely (Lamme, 2003; Dehaene, Changeux, Naccache, Sackur, & Sergent, 2006). In an influential paper, Lamme (2003) has argued against some traditional views which equated awareness with the focus of attention. Instead, he proposed that we consider attention and awareness separate. In this view, multiple inputs reach a conscious state but only the attended ones are reportable. He presented evidence form a change detection paradigm, where several items are presented on a display. After a sufficient delay, the display is shown again but with one item changed. If the participants are now asked to judge whether an item has changed or not, their performance will indicate that only about 4 items have been remembered, which is the limit of working memory (Cowan, 2001). However, if one of the locations is cued during the interval between the presentation of the two displays, performance goes up significantly (Sligte, Scholte, & Lamme, 2008; Sligte, 2010; Pinto, Sligte, Shapiro, & Lamme, 2013). This effect is not due to an after-image (Sligte et al., 2008) but rather shows that by shifting attention towards the cued location, a consciously processed item can become reportable. Figure 2 visualizes the relation between attention and awareness proposed by Lamme (2003).
Figure 2 The distinction between conscious and attended item, by Lamme (2003).
Similarly, Dehaene and colleagues (2006), developed a taxonomy which places top-down attention and bottom-up stimulus strength on two axes which jointly contribute to conscious awareness. This allows for distinguishing between four different types of processing: (1) Unattended subliminal processing happens when the stimulus strength is weak and attention is absent. The stimulus evokes little neuronal activation and little to no priming. (2) Attended subliminal processing is associated with weak stimulus information which receives sufficient top-down attention. Such stimuli are not reportable but can induce short-lived priming effects and cause strong feedforward activity in the brain. (3) Preconscious processing happens with sufficiently strong stimuli which are not attended. Such stimuli are processed in both a feedforward and recurrent manner but the activity is confined to local areas. On a behavioural level, preconscious stimuli induce priming at multiple levels but are not reportable. (4) Conscious processing is the effect of top-down attention focused on stimuli with sufficient strength. Conscious stimuli induce global, recurrent neuronal activity and durable, reportable information.
Figure 3 The taxonomy developed by Dehaene et al. (2006) views attention as just one of two factors contributing to visual awareness, the other being stimulus strength.
Both of these highly influential accounts in consciousness research posit a distinction between attention and awareness. In our model, we regard attention as necessary but not sufficient to produce visual awareness, as theorised by Lamme (2003). We incorporate the two factors underlining visual awareness, as described by Dehaene et al (2006): an attentional, top-down factor and a stimulus-driven bottom-up factor. We claim that the attentional component works in a dichotomous manner, distinguishing between attended items characterized by sufficient activation strength, and unattended items for which neural activity is below the threshold. We claim that all attended items can potentially become conscious, and that the extent of that conscious awareness depends on the stimulus strength.
According to our hypothesis then, in experiments which manipulate attention, such as the attentional blink, participants should not report intermediate conscious states. On the other hand, in experiments manipulating stimulus strength, such as backward masking, participant experience those intermediate states. Therefore, we hypothesize that the choice of the experimental paradigm influences whether intermediate conscious states are observed and, thus, affects whether awareness is concluded to be continuous or all-or-none.
Testing the model
This experiment examines the predictions of the two-process model. We adapted a procedure previously used by Asplund and colleagues (2014), who used continuous stimuli to measure visual awareness in a colour-detection task, originally developed by Bays and colleagues (2009). In this paradigm, participants have to report the colour they had seen by selecting a value on a colour wheel. This allows to calculate the response error in continuous terms as a degree difference between the position of the target colour and the selected colour. The benefit of this approach is that it produces a distribution of errors on a continuous scale, which can then be modelled as a mixture of unseen (guess) trials, taken from a uniform distribution, and seen trials, taken from a circular normal distribution (i.e. Von Mises distribution) centred on the target value with a certain precision. This is ideal for answering the question at hand, as it allows to test whether the resulting distribution can be modelled better as a mixture of seen trials with fixed precision (all) and uniformly distributed guesses (none), or as a continuum of trials with varying degree of precision.
Using this paradigm in combination with attentional blink, Asplund et al. (2014) presented a visual stream consisting of circles (distractors) and squares (targets). All stimuli had a colour taken from an equiluminant colour wheel, except for the first target which was either black or white. Participants were instructed to pay attention to the colour of the squares and ignore the circles. The second target appeared 1, 2, 4 or 8 serial positions after the first. The on-screen duration of each stimulus was not fixed but rather adjusted throughout the experiment to maintain around 60% performance, averaging at 115 ms, SD = 10 ms. For each interval and subject, the researchers computed encoding probability and precision using mixture modelling (Bays et al., 2009). Encoding probability was the percentage of non-guess trials, while precision was the inverse of the width of the component circular normal distribution, reflecting the quality of the encoded percept. They found that encoding probability was dropping with a decrease in interval, reflecting an increasing guess rate. Precision, however, was unaffected by the manipulation. The researchers concluded that conscious perception is all-or-none.
Here, we replicated the experiment of Asplund et al. (2014) in combination with backward masking. We predicted that this modification would allow us to observe states of graded awareness, by tapping into the bottom-up process. Therefore, we expected response precision to be affected by stimulus quality, and encoding probability to be unaffected or affected to a lesser degree.
Methods
Procedure
Following Asplund and colleagues (2014), we used a colour-detection task with continuous stimuli in a backward masking task where we manipulated stimulus strength. Participants had to choose the colour they had seen by selecting a value on a colour wheel and we examined the error distribution of their responses. When the stimulus is consciously perceived, the error will be distributed around the actual value with precision dependent on the quality of the representation. If the stimulus is not consciously perceived, the response will be random so the measured error will be distributed uniformly. The observed error distribution can thus be modelled as a mixture of the two component distributions. This allowed us to measure the probability of perceiving the stimulus and the quality of the percept.
If participants responded in an all-or-none manner, then manipulating the stimulus strength should affect the probability of perceiving the stimulus but not the precision of the response. If, on the other hand, they responded continuously, then an increase in stimulus strength should lead to an increase in response precision.
16 participants completed the study (8 male, mean age = 35.06, SD = 8.31). Participants were recruited through an online recruitment system from the Humboldt University of Berlin and were payed 10.50€ in compensation. One participant was excluded due to overall poor performance. The task consisted of a centrally-presented square (1.2° × 1.2°; all values in visual angles) on a grey background (#787878), followed by a mask. The participants then had to navigate along a colour wheel to choose the perceived colour. Afterwards, they were presented with a confidence rating, where they had to press a number corresponding to the confidence in their response on a scale 1-4. The stimuli were one of 180 equiluminant colours (drawn from the CIEL*a*b* colour space with D50 white point, centred at L = 54, a = 18, b = 12, with a radius of 59). The colours were calibrated with a spectrometer and JetiVal software to minimize the distortions caused by the monitor used for testing. The resulting colour palette had an average luminance of 57.56 and standard deviation of 0.84. The average distortion of a and b values, i.e. the average difference between the desired and obtained values was 0.74 and 1.26, with standard deviations of 0.31 and 0.51, respectively. The stimulus was presented for 32ms. The mask was a centrally-presented square (3.5° × 3.5°), composed of 169 smaller squares chosen randomly from the set of 180 colours used in the experiment. The squares composing the mask were scrambled randomly and each had a unique colour. The mask was presented for 500ms.
Figure 5 The colour wheel used in the experiment. After giving a response, participants were given feedback showing the actual colour (green line) in comparison with their response (black line).
We manipulated the interval between the stimulus and the mask, ranging from 0ms to 64ms in four steps: 0ms, 16ms, 32ms and 64ms (henceforth referred to as intervals 1, 2, 3 and 4, respectively). The colour wheel was a ring with a radius of 5.8° and the width of 1.2°. All stimuli were presented on a 19” DellLCD monitor (model 1908FPc) with a resolution 1280×1024 and refresh rate 60Hz using Matlab 2011b and Psychtoolbox. Participants were seated on a leather armchair at the distance of 57cm in front of the screen (total viewing angle 33.7° × 27.8°). Participants selected their response by moving a line adjacent to the colour wheel in a clockwise or anti-clockwise direction by pressing one of the two corresponding buttons. On each trial, the starting position of the response line was randomized. Participants performed 8 blocks of 60 trials each in one experimental session, with a 10 minute break after the fourth block. Before the experiment, participants performed 25 training trials. The total duration of the experiment was 90 minutes and included 480 experimental trials per participant.
Analysis
Response error was calculated for each trial as the difference between the participant’s response and the presented colour. Errors were modelled as a weighted mixture of two distributions, corresponding to perceived and unperceived trials. Perceived responses were drawn from a circular normal distribution defined by its mean (µ, a measure of bias) and concentration (k, a measure of spread, converted to σ) and unperceived responses were drawn from a uniform distribution. The mixture parameter Pe (relative weight of the circular normal distribution) reflected the probability of perceiving the stimulus. The parameter values for each variable and stimulus-mask interval were computed using maximum likelihood estimation.
Furthermore, for each trial we calculated the probability of the response originating from the circular normal distribution, i.e. the probability that the target was encoded in that trial. The analysis code was developed by Bays and colleagues (2009; Schneegans & Bays, 2016). All analyses were performed using Matlab R2015a and JASP 0.8.1.2 (JASP Team, 2016).
Results
We assessed the probability of encoding the target and the width of the Von Mises distribution corresponding to the target response. All tests were performed with repeated-measures analyses of variance (ANOVA) with Bonferroni-corrected post-hoc tests, and Bayesian paired-samples planned comparisons. The significance level was set at 0.05. The Bayes factor for planned comparisons is reported next to post-hoc test pairs. The Mauchly’s test was used to test for data sphericity. Wherein the test indicated violation of the assumption of sphericity, the Greenhouse-Geisser correction was applied to the ANOVA. F values using the correction are denoted with a subscript (FGG). In the post-hoc comparison tables below, rows with significant p values are marked in bold, and rows with values showing a trend, which did not reach the significance threshold, are underlined.
First, we tested whether our manipulation was successful by looking at the differences in response error by interval. The assumption of sphericity was violated (W = 0.267, p = 0.005). ANOVA revealed a significant effect of interval on response error, FGG(25.060,1.790) = 195.8, p < 0.001, η2 = 0.933. For each interval pair, the response error was lower at a longer interval, a finding typical of backward masking paradigm. The average errors (in degrees) for each interval are presented in Table 1 and Figure 6. Table 2 contains the corresponding t values, along with the Bayes factor for planned comparisons.
Figure 6 Response error by interval.
interval Mean SD
0 ms 71.20 8.211
16 ms 49.64 15.234 32 ms 30.29 14.550
64 ms 17.05 6.913
Mean Difference SE t p bonf BF10 0 ms 16 ms 21.56 2.736 7.879 < .001 22454.10 32 ms 40.91 2.753 14.862 < .001 3.178 × 107 64 ms 54.15 1.426 37.965 < .001 4.382 × 1012 16 ms 32 ms 19.35 1.744 11.097 < .001 987042.80 64 ms 32.59 2.882 11.311 < .001 1.231 × 106 32 ms 64 ms 13.24 2.384 5.554 < .001 774.80
Table 2 Post-hoc comparisons for response error by interval.
Second, we calculated response bias to check whether participants were consistently deviating from the target value. The average response bias for each interval was -0.034, -0.007, 0.013, and 0.005 for intervals 1, 2, 3 and 4, respectively (all values in radians). ANOVA did not reveal a significant effect of interval on response bias, FGG(1.648,23.077) = 0.537, p = 0.558, η2= 0.037. The Bayesian ANOVA favoured the null model, BF10 = 0.15. Post-hoc comparisons did not reveal significant differences between any interval pairs, see Table 3 for corresponding values.
Mean Difference SE t p bonf BF10
0 ms 16 ms -0.027 0.039 -0.678 1 0.321 32 ms -0.047 0.041 -1.143 1 0.457 64 ms -0.039 0.059 -0.667 1 0.319 16 ms 32 ms -0.02 0.025 -0.806 1 0.348 64 ms -0.013 0.036 -0.35 1 0.277 32 ms 64 ms 0.007 0.028 0.262 1 0.270
Table 3 Post-hoc comparisons for bias by interval.
Encoding probability
The assumption of sphericity was violated (W = 0.14, p < 0.001). ANOVA revealed that the probability of encoding the target (Pe) was strongly affected by interval, FGG(1.538,21.532) = 16.73, p < 0.001, η2 = 0.544. Post-hoc comparisons using the Bonferroni correction revealed significant differences between interval pairs: 1 and 4 (t(14) = -5.79, p < 0.001, BF10 = 1115.49), 2 and 4 (t(14) = -7.03, p < 0.001, BF10 = 6964.33), as well as 3 and 4 (t(14) = -4.89, p = 0.001, BF10 = 266.51). A trend was observed in the differences between intervals 1 and 3 (t(14) = -2.79, p = 0.086, BF10 = 8.18), as well as 2 and 3 (t(14) = -2.86, p = 0.076, BF10 = 9.11) but the results did not reach statistical significance.
Figure 7 Encoding probability by interval. Encoding probability is the contribution of the Von Mises (circular normal) distribution to the observed results.
Mean Difference SE t p bonf BF10
0 ms 16 ms -0.061 0.042 -1.462 0.996 1.15 32 ms -0.181 0.065 -2.792 0.086 8.18 64 ms -0.287 0.050 -5.788 < .001 1115.49 16 ms 32 ms -0.120 0.042 -2.859 0.076 9.12 64 ms -0.226 0.032 -7.025 < .001 6964.33 32 ms 64 ms -0.106 0.022 -4.888 0.001 266.51
Table 4 Post-hoc comparisons for target encoding probability by interval.
Precision
Precision was operationalized as the inverse of the width of the underlying Von Mises distribution, corresponding to the target response. The measure of width was expressed in terms of standard deviation. The assumption of sphericity was violated (W = 0.10, p < 0.001). ANOVA revealed that width was strongly affected by interval, FGG(1.826,25.561) = 28.90, p < 0.001, η2 = 0.674. Post-hoc comparisons using the Bonferroni correction revealed that response precision was lower at interval 1 than at intervals 2, 3 and 4 (t(14) = 5.34, p < 0.001, BF10 = 274.38; t(14) = 6.17, p < 0.001, BF10 = 993.85; and t(14) = 7.13, p < 0.001, BF10 = 4026.71, respectively). Response precision was also higher at interval 4 than at interval 3 (t(14) = 4.27, p = 0.005, BF10 = 47.64). Furthermore, precision was lower at interval 2 than at interval 4 (t(14) = 3.05, p = 0.052, BF10 = 6.29) but the difference was marginally below the threshold of statistical significance.
Figure 8 Width of the Von Mises distribution by interval.
Mean Difference SE t p bonf BF10
0 ms 16 ms 0.305 0.057 5.336 < .001 274.384 32 ms 0.421 0.068 6.167 < .001 993.851 64 ms 0.479 0.067 7.128 < .001 4026.708 16 ms 32 ms 0.117 0.055 2.117 0.316 1.478 64 ms 0.175 0.057 3.052 0.052 6.294 32 ms 64 ms 0.058 0.014 4.265 0.005 47.644
Table 5 Post-hoc comparisons for response width by interval.
Confidence ratings
The assumption of sphericity was violated (W = 0.037, p < 0.001). ANOVA revealed that response confidence was strongly affected by interval, FGG(1.191,16.673) = 34.87, p < 0.001, η2 = 0.714. Post-hoc comparisons using the Bonferroni correction revealed significant differences between each pair of intervals. As the interval increased, the confidence rating increased as well. The corresponding t values are presented in Table 6.
Figure 9 Confidence rating by interval
Mean Difference SE t p bonf BF10
0 ms 16 ms -0.379 0.082 -4.629 0.002 174.3 32 ms -0.804 0.154 -5.210 < .001 448.6 64 ms -1.273 0.203 -6.278 < .001 2350.3 16 ms 32 ms -0.425 0.086 -4.949 0.001 294.2 64 ms -0.894 0.139 -6.443 < .001 3004.2 32 ms 64 ms -0.469 0.073 -6.455 < .001 3057.1
Table 6 Post-hoc comparisons for confidence by interval.
To exclude the possibility that the differences in confidence ratings were driven by low confidence on guessing trials, a subsequent analysis included only the trials on which the probability of encoding the target was higher than 0.5. The assumption of sphericity was violated (W = 0.081, p < 0.001). ANOVA revealed that response confidence was strongly affected by interval, FGG(1.299,18.187) = 36.93, p < 0.001, η2 = 0.725. Post-hoc comparisons using the Bonferroni correction revealed significant differences between each pair of intervals. As the interval increased, the confidence rating increased as well. The corresponding t values are presented in Table 7.
Figure 10 Confidence rating by interval. Guessing trials—i.e. trials on which the estimated encoding probability was equal to or lower than 0.5—were excluded from the analysis.
Mean Difference SE t p bonf BF10
0 ms 16 ms -0.444 0.091 -4.899 0.001 135.70 32 ms -0.807 0.148 -5.461 < .001 334.63 64 ms -1.230 0.184 -6.679 < .001 2121.87 16 ms 32 ms -0.363 0.079 -4.612 0.002 84.79 64 ms -0.786 0.121 -6.502 < .001 1638.54 32 ms 64 ms -0.423 0.066 -6.433 < .001 1479.96
Table 7 Post-hoc comparisons for confidence by interval. Guessing trials—i.e. trials on which the estimated encoding probability was equal to or lower than 0.5—were excluded from the analysis.
Finally, we plotted the confidence ratings distribution for each interval. We did not observe a dual-cluster structure reported by Sergent and Dehaene (2004) and Del Cul et al. (2007). Rather, confidence ratings peaked at a single value or at two adjacent values: value 1 for 0 ms (38% of trials), values 2 and 3 for 16 ms (28% and 30%, respectively), value 3 for 32 ms (37%) and value 4 for 64 ms (50%). Repeated-measures ANOVAs with confidence rating as factor revealed a significant effect of confidence at intervals 1 (FGG(26.098,1.864) = 4.61, p = 0.021, η2 = 0.248), 3 (FGG(20.745,1.482) = 7.31, p = 0.007, η2 = 0.343), and 4 (FGG(17.074,1.220) = 15.32, p < 0.001 η2 = 0.522). Post-hoc comparisons were significant only at interval 3 between confidence ratings 1-2 (t(14) = -3.650, p = 0.016), 1-3 (t(14) = -6.652, p < 0.001), and 2-3 (t(14) = -5.732, p < 0.001), as well as at interval 4 between all confidence ratings except 3-4 (see Table 8). Figure 11 presents the distribution of confidence ratings by interval.
Figure 11 Distribution of confidence ratings by interval.
Confidence Mean Difference SE t p bonf
1 2 -0.085 0.028 -3.054 0.051 3 -0.321 0.052 -6.118 < .001 4 -0.466 0.076 -6.136 < .001 2 3 -0.236 0.048 -4.892 0.001 4 -0.381 0.097 -3.945 0.009 3 4 -0.145 0.122 -1.190 1.000
Table 8 Post-hoc comparisons for confidence ratings at interval 4 (64 ms).
Discussion
We measured how the target-mask interval in the backward masking paradigm affects target encoding probability and response precision. Unlike Asplund and colleagues (2014), we found that precision was affected by our manipulation. Within our model, higher performance at interval 2 compared to interval 1 (16 ms and 0 ms, respectively) can only be explained by increased precision. Although the picture is not entirely clear when comparing longer intervals, we conclude that at the shortest intervals, participants were perceiving the target in a graded manner and that a 0 ms interval invoked less precise visual percepts than a 16 ms interval. Quite possibly, given the simple nature of the stimuli used, a 16ms interval is sufficient to build a high quality representation thereof, and a further increase in interval duration contributes to only a minor enhancement of the percept. That would explain the significant difference between 32 and 64 ms but not between 16 and 32 ms – a 32 ms difference in the interval duration (as in the former) leads to a measureable perceptual difference, whereas a 16 ms difference (as in the latter) does not. Additionally, it is quite likely that there is no simple linear relation between the interval duration and the quality of the perceptual representation.
This claim is further corroborated by the observation that confidence increases linearly with the increase in interval. This effect cannot be explained by the decreasing number of low-confidence guess trials, since the effect persists when trials with low encoding probability are excluded.
Moreover, the distribution of confidence rating responses suggests that participants were using the confidence scale in a continuous manner, unlike in the AB experiments conducted by Sergent and Dehaene (2004) or Del Cul et al. (2006, 2007).
In our experiment, both encoding probability and response precision were affected by the manipulation, with precision being the only factor explaining differences at lower intervals. This means that, even though the guess rate was decreasing with the increase in interval (at least at the longer intervals), response precision was increasing as the interval was getting longer (at least at the shorter intervals, but also between the two longest intervals). This result suggests that targets were perceived in a continuous manner. Given that our experiment closely followed the design of Asplund et al. (2014), the results suggest that the manipulation of the backward masking paradigm is affecting a different mechanism than that of the attentional blink. This difference is captured in our two-process model of visual awareness in that AB acts on a top-down, attention-dependent two-process with all-or-none characteristics, whereas backward masking involves the manipulation of a bottom-up stimulus-driven process which is continuous.
Alternative explanations for the observed results must be considered. Even at the lowest intervals, participants were performing fairly well—average error was 71°, whereas the maximum error that could be made in this task is 180°. Furthermore, mixture modelling revealed that encoding probability was 0.62 at the lowest interval. Given this, one might reason that there is in fact a threshold for stimulus quality, below which no awareness occurs, and above which awareness is increasing gradually. Ramsøy and Overgaard (2004) pointed to this problem, noticing a ceiling effect in their PAS scores—participants rarely reported having no experience of the stimulus at all. The alternative explanation for our results, then, could be that we only sampled from the stimulus quality spectrum which is above that threshold and, were the task to be made more difficult, we would observe a non-continuous increase in visual awareness at values near the threshold. This argument cannot be refuted using our setup, as the durations used (32 ms target on-screen duration at the shortest interval) were close to the technical limitations of the equipment used. Perhaps the simple nature of the stimuli (colours) made the task easy. However, using more complex—and, thus, more difficult to perceive—stimuli would make it more challenging to create a continuous scale on which the stimuli could be placed in order to be able to capture a continuous measure of error.
Nevertheless, we can refute this argument by analysing it from a theoretical point. Here, we were concerned with whether visual awareness is continuous or all-or-none. By showing that manipulating stimulus quality leads to continuous changes in stimulus awareness, we’ve provided an argument against the all-or-none view. Whether or not there exists a threshold is irrelevant: assuming we are sampling values above that threshold, if visual awareness were in fact all-or-none, we would not find a continuous increase—instead, all perceived trials would have the same, maximum precision. Asplund and colleagues (2014) concluded that “conscious perception, at least at central stages of information processing, is all-or-none” (p. 829). This is clearly not the case here.
In conclusion, the two-process model captures well the results obtained in attentional blink and backward masking experiments. Further research is required to see whether it generalises to other paradigms used in consciousness research, such as inattentional blindness or continuous flash suppression. This paper shows that no single paradigm is sufficient to make far-reaching claims about the nature of consciousness. Consciousness is a complex phenomenon—arguably the most complex
one in cognitive sciences—and as such, multiple approaches to studying it must be integrated before one concludes that it is either continuous or all-or-none.
References
Asplund, C. L., Fougnie, D., Zughni, S., Martin, J. W., & Marois, R. (2014). The Attentional Blink Reveals the Probabilistic Nature of Discrete Conscious Perception. Psychological Science, 25(3), 824– 831. https://doi.org/10.1177/0956797613513810
Bar, M., Tootell, R., Schacter, D., Greve, D., & Fischl, B. (2001). Cortical mechanisms specific to explicit visual object recognition. Neuron. Retrieved from
http://www.sciencedirect.com/science/article/pii/S0896627301002240
Bays, P. M., Catalao, R. F. G., Husain, M., J., D., C., K., & M., H. (2009). The precision of visual working memory is set by allocation of a shared resource. Journal of Vision, 9(10), 7–7.
https://doi.org/10.1167/9.10.7
Breitmeyer, B., & Ogmen, H. (2000). Recent models and findings in visual backward masking: A comparison, review, and update. Perception & Psychophysics. Retrieved from
http://link.springer.com/article/10.3758/BF03212157
Broadbent, D., & Broadbent, M. (1987). From detection to identification: Response to multiple targets in rapid serial visual presentation. Attention, Perception, & Psychophysics. Retrieved from http://www.springerlink.com/index/B66G94N882K14378.pdf
Christensen, M. S., Ramsøy, T. Z., Lund, T. E., Madsen, K. H., & Rowe, J. B. (2006). An fMRI study of the neural correlates of graded visual perception. NeuroImage, 31(4), 1711–1725.
https://doi.org/10.1016/j.neuroimage.2006.02.023
Cowan, N. (2001). The magical number 4 in short-term memory: A reconsideration of mental storage capacity. Behavioral and Brain Sciences, 24(1), 87–114.
https://doi.org/10.1017/S0140525X01003922
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. https://doi.org/10.1016/j.tics.2006.03.007
Del Cul, A., Baillet, S., Dehaene, S., Breitmeyer, B., Cul, A. Del, Dehaene, S., … Rensink, R. (2007). Brain Dynamics Underlying the Nonlinear Threshold for Access to Consciousness. PLoS Biology,
5(10), e260. https://doi.org/10.1371/journal.pbio.0050260
Del Cul, A., Dehaene, S., & Leboyer, M. (2006). Preserved subliminal processing and impaired conscious access in schizophrenia. Archives of General Psychiatry, 63(12), 1313–1323. https://doi.org/10.1001/archpsyc.63.12.1313
Grill-Spector, K., Kourtzi, Z., & Kanwisher, N. (2001). The lateral occipital complex and its role in object recognition. Vision Research. Retrieved from
http://www.sciencedirect.com/science/article/pii/S0042698901000736 JASP Team. (2016). JASP (Version 0.8.1.2). Computer Software.
Kanwisher, N. (2001). Neural events and perceptual awareness. Cognition. Retrieved from http://www.sciencedirect.com/science/article/pii/S0010027700001256
Lamme, V. A. F. (2003). Why visual attention and awareness are different. Trends in Cognitive
Sciences, 7(1), 12–18. Retrieved from
http://www.sciencedirect.com/science/article/pii/S136466130200013X
Nieuwenhuis, S., & de Kleijn, R. (2011). Consciousness of targets during the attentional blink: a gradual or all-or-none dimension? Attention, Perception, & Psychophysics, 73(2), 364–373.
https://doi.org/10.3758/s13414-010-0026-1
Overgaard, M., Rote, J., Mouridsen, K., & Ramsøy, T. Z. (2006). Is conscious perception gradual or dichotomous? A comparison of report methodologies during a visual task.
https://doi.org/10.1016/j.concog.2006.04.002
Pinto, Y., Sligte, I. G., Shapiro, K. L., & Lamme, V. a. (2013). Fragile visual short-term memory is an object-based and location-specific store. Psychonomic Bulletin & Review, 20(4), 732–739. https://doi.org/10.3758/s13423-013-0393-4
Pretorius, H., Tredoux, C., & Malcolm-Smith, S. (2016). Subjective awareness scale length influences the prevalence, not the presence, of graded conscious states. Consciousness and Cognition, 45, 47–59. https://doi.org/10.1016/j.concog.2016.08.007
Ramsøy, T. Z., & Overgaard, M. (2004). Introspection and subliminal perception. Phenomenology and
the Cognitive Sciences, 3(1), 1–23. https://doi.org/10.1023/B:PHEN.0000041900.30172.e8
Sandberg, K., Bibby, B. M., Timmermans, B., Cleeremans, A., & Overgaard, M. (2011). Measuring consciousness: Task accuracy and awareness as sigmoid functions of stimulus duration.
Consciousness and Cognition, 20(4), 1659–1675. https://doi.org/10.1016/j.concog.2011.09.002
Sandberg, K., Timmermans, B., Overgaard, M., & Cleeremans, A. (2010). Measuring consciousness: Is one measure better than the other? Consciousness and Cognition, 19(4), 1069–1078.
https://doi.org/10.1016/j.concog.2009.12.013
Schneegans, S., & Bays, P. M. (2016). No fixed item limit in visuospatial working memory. CORTEX, 83, 181–193. https://doi.org/10.1016/j.cortex.2016.07.021
Sergent, C., & Dehaene, S. (2004). Is consciousness a gradual phenomenon? Evidence for an all-or-none bifurcation during the attentional blink. Psychological Science, 15(11), 720–728. https://doi.org/10.1111/j.0956-7976.2004.00748.x
Sklar, A. Y., Levy, N., Goldstein, A., Mandel, R., Maril, A., & Hassin, R. R. (2012). Reading and doing arithmetic nonconsciously. Proceedings of the National Academy of Sciences, 109(48), 19614– 19619. Retrieved from http://www.pnas.org/content/109/48/19614.short
Sligte, I. G. (2010). Detailed sensory memory, sloppy working memory. Frontiers in Psychology, 1. https://doi.org/10.3389/fpsyg.2010.00175
Sligte, I. G., Scholte, H. S., & Lamme, V. A. F. (2008). Are There Multiple Visual Short-Term Memory Stores? PLoS ONE, 3(2), e1699. https://doi.org/10.1371/journal.pone.0001699
Tulving, E., & Schacter, D. (1990). Priming and human memory systems. Science. Retrieved from http://www.jstor.org/stable/pdf/2873625.pdf
