Conscious access in the near absence of attention: critical extensions on the dual-task paradigm

Abstract

Whether conscious perception requires attention remains a topic of intense debate. While certain complex stimuli such as faces and animals can be discriminated outside the focus of spatial attention, many simpler stimuli cannot. Because such evidence was obtained in dual-task paradigms involving no measure of subjective insight, it remains unclear whether accurate discrimination of unattended complex stimuli is the product of automatic, unconscious processing, as in blindsight, or is accessible to consciousness. Furthermore, these paradigms typically require extensive training over many hours, bringing into question whether this phenomenon can be achieved in naive subjects. We developed a novel dual-task paradigm incorporating confidence ratings to calculate metacognition and adaptive staircase procedures to reduce training. With minimal training, subjects were able to discriminate face-gender in the near absence of top–down attentional amplification, while also displaying above-chance metacognitive accuracy. By contrast, the discrimination of simple coloured discs was significantly impaired and metacognitive accuracy dropped to chance-level, even in a partial-report condition. In a final experiment, we used blended face/disc stimuli and confirmed that face-gender but not colour orientation can be discriminated in the dual task. Our results show direct evidence for metacognitive conscious access in the near absence of attention for complex, but not simple, stimuli.

This article is part of the theme issue ‘Perceptual consciousness and cognitive access'.

1. Introduction

The perplexing co-dependency between attention and consciousness has been the subject of philosophical and scientific debate for well over a century [1]. One feature of this debate that has risen to prominence in recent years concerns the necessity of top–down attentional amplification for conscious perception (the necessity claim) [2–4]. While several noteworthy theories of consciousness remain divided on this claim, scientific enquiry has made progress in its attempts to independently manipulate top–down attention and visual consciousness using a variety of tasks and illusions¹ [5,17,18].

The relationship between visual consciousness and top–down attentional amplification has been primarily investigated with the dual-task paradigm [19–21]. In this paradigm, a subject's attention is spatially drawn to a very demanding central task at the same time as a secondary stimulus is briefly presented in the periphery. If performance on the peripheral task in this diverted attention condition is identical to that without the central task, top–down attention is claimed unnecessary for the peripheral task. This paradigm has been employed to examine the requirement of top–down attention for discriminating many categories of stimuli. For example, a simple, low-level visual distinction such as discriminating the orientation of a coloured shape can be 80% correct under full attention and yet fall to chance (50% correct) when attention is diverted, demonstrating a necessity of spatial attention [22,23].

Remarkably, however, performance for certain visual discriminations that seem intuitively more complex such as categorizing face-gender or classifying whether natural scenes include animals does not differ between the full and diverted attention conditions [23–25]. These results suggest that for certain complex stimuli, attention may not be necessary for conscious perception, a result taken as empirical evidence for a dissociation between top–down attention and consciousness [5,26].

However, dual-task studies investigating such a dissociation are confronted with various criticisms that question this conclusion. Among these, we address the most critical four in our experiments. First, dual-task experiments typically employ extensive training of thousands of trials. In the case of Reddy et al. [23], training took between 6 and 12 h before subjects achieved proficiency. This involved as many as 5760 training trials, more than the experiment itself. Training is known to influence the attentional requirements of the dual task [27–30] and other paradigms including the attentional blink [31,32]. Specifically, the extent to which tasks are influenced by inattention differs between highly trained and naive subjects, which poses critical limitations on the conclusions that can be drawn when such tasks are employed to examine the relationship between attention and consciousness [28–30].

The second challenge concerns the nature of the control experiments employed in the studies listed above [23–25]. In these experiments and under the dual-task condition, certain categories of stimuli such as rotated letters or bisected discs are shown to be impossible to discriminate at above-chance levels [17]. A typical account of this scenario, inattentional blindness [33], suggests that without top–down attentional amplification, these stimuli fail to reach consciousness. An alternative account is that the inability to discriminate such items is not perceptual but results from response interference. That is, without top–down attention, while subjects are responding to the central task, they forget the stimulus because peripheral representations decay very quickly. In other words, the standard dual-task paradigm cannot exclude a possibility of conscious visibility with rapid forgetting of those stimuli that result in chance-level performance under the diverted attention condition [34–37].

The third criticism concerns the difference in attentional draw between different categories of stimuli. Of particular concern for the studies by Reddy et al. [23,25], faces are known to strongly attract (bottom–up) attention, possibly due to ecological importance and the presence of dedicated neural resources [38–40]. They capture attention in individuals as young as six weeks old [41], and they impair processing of visual objects that are presented elsewhere at the same time [42,43]. Perhaps this attraction to faces accounts for why this category of stimulus can be discriminated in the dual task while less salient items, such as bisected discs, are missed?

Finally, it remains unclear whether successful dual-task performance for categories such as face-gender is in fact conducted using information accessible to consciousness. Recent literature has accumulated considerable evidence of above-chance behavioural performance that is not accompanied with consciousness, such as blindsight [44–46]. Non-conscious stimulus processing has been observed not only for simple discriminations, but also for complex, high-level stimuli such as upright faces [47–50] (for reviews, see [51,52]). Thus, achieving highly accurate dual-task performance in tasks with complex stimuli such as faces does not guarantee that these discriminations are performed consciously [4]. In fact, in combination with excessive training, there is reason to suspect that at least some aspects of successful discrimination under dual-task conditions may be a product of unconscious processing [49].

In this paper, we addressed these four criticisms by substantially improving the dual-task paradigm in four ways. First, we employed an adaptive staircase procedure [53] to reduce the amount of training typical of dual-task studies [23,24]. Second, our partial-report condition made both the central and peripheral targets task-relevant but required subjects make only one response per trial, removing the concern that peripheral target representations decay beyond reportability while subjects respond to the central task. Third, we accounted for faces attracting bottom–up spatial attention by blending face and disc stimuli through transparency (alpha (α) blending) and examining whether the colour orientation of the disc is reportable when the stimuli are co-located. Fourth, we directly assessed metacognitive insight for unattended peripheral stimuli, a signifier of conscious access, by quantifying trial-by-trial confidence ratings and perceptual awareness judgements as a function of task accuracy [54–57]. With these improvements over previous approaches, we critically assessed the necessity of top–down attention for conscious access of stimuli with high and low complexity.

2. General methods

(a). Subjects

Twenty-four subjects participated in our study, eight for each of our three experiments. Participant numbers were determined from those studies that employed the dual-task paradigm: typically between four and eight [23–25,58]. A power analysis (power of 0.8, assumed correlation of 0.5, one-tailed t-test) based on observations from the disc task in Reddy et al. [23] revealed that a sample size of 3 would be sufficient to find a difference comparable to or larger than their study. Subjects were recruited from the student and staff bodies of Monash University and were paid for their involvement in the study. All had normal or corrected-to-normal vision and provided informed written consent in accordance with the guidelines of the Monash University Human Research Ethics Committee and the recommendations of the Declaration of Helsinki.

(b). Apparatus

All experiments were performed on a MacBook Pro laptop connected to a 22-inch SMI monitor approximately 60 cm from the subject. Refresh rate of the monitor was fixed at 60 Hz with 1680 × 1050 pixels screen resolution. The experiments were programmed and conducted using the Psychophysics toolbox extension for Matlab.

(c). Stimuli

(i). Central letter discrimination

The central stimulus for all experiments was a cluster of five uppercase characters presented in white, Helvetica script at 35 pixels text height (approx. 1° visual angle on our set-up), each rotated at a randomly selected angle. The coordinates of these five letters were fixed, one presented centrally at fixation and the remaining four located directly above, below and on both sides of this point, approximately 3° from fixation. These five letters were either all the same (all ‘T’ or all ‘L’) or contained a single differing character (i.e. one ‘T’ among four ‘L's and vice versa). An uppercase letter ‘F’ individually masked each character for the remainder of the trial following a short, temporal delay that was adjusted to achieve 70% discrimination accuracy across central, single-task blocks (see quick estimate of threshold (QUEST) staircase procedure, §2e). This letter discrimination task has proven effective in maintaining the focus of attention at the fixational point, leaving little spatial attention available at the periphery [22–25].

(ii). Peripheral discrimination

The peripheral stimulus categories consisted of faces (experiment 1), discs (experiment 2) or blended face/discs (experiment 3) (figure 1a). On each trial, one such stimulus subtending 2.5° of visual angle was displayed at the periphery. This peripheral stimulus was randomly positioned at one of four locations centred on the corners of an imaginary rectangle 8° × 10° of visual angle in dimensions. After a short temporal delay, a mask replaced the peripheral stimulus (figure 1a). This delay was adjusted such that single-task discrimination was held at 70% accuracy (see QUEST staircase procedure, §2e, and figure 1b).

Figure 1.

Stimuli and task structure. (a) Stimuli used for experiment 1: gender discrimination, experiment 2: bisected disc discrimination, and experiment 3: blended face/disc discrimination. (b) Trial sequence. After a variable period of fixation (200–400 ms), five randomly rotated letters (Ls and/or Ts) are presented in the centre. After one frame (16.7 ms), a peripheral target (here, a female face) appeared in the periphery. Following a short delay, or stimulus onset asynchrony (SOA), both central and peripheral targets were masked. In blocks involving the central task (the central single- and dual-task conditions), subjects reported on the letter stimuli at the centre of the display. Subjects made an eight-alternative-forced-choice (8AFC) response with a single mouse click, signalling their discrimination (‘S’ for same (all Ts or all Ls) and ‘D’ for different (T among Ls or L among Ts)) and four-level subjective rating (confidence in experiments 1 and 2, perceptual awareness in experiment 3). In blocks involving the peripheral task (the peripheral single- and dual-task conditions), subjects reported on the peripheral stimulus. Again, a single mouse click was used for an 8AFC decision on face-gender (experiments 1 and 3: M, male; F, female) or disc colour-orientation (experiments 2 and 3: a picture of a red–green disc and a green–red disc). (c) Summary of partial- and whole-report conditions in experiment 2. During dual-task blocks, task relevance remained consistent but depending on the report condition, either one or two responses were required.

(d). Procedure

Dual-task experiments contrast performance in the single-task condition, where a central or a peripheral task is conducted in isolation, against that in the dual-task condition, where both the central and peripheral stimuli are task-relevant. The physical appearance of the experiment should be identical across these conditions, with the only difference being the task relevance of the stimuli (figure 1c). Written instructions at the beginning of each block informed subjects which task was required.

(i). Single-task conditions

In the central and peripheral single-task conditions, subjects were presented with one response screen and made a single eight-alternative-forced-choice (8AFC) report per trial (figure 1b: response panels enclosed in green and yellow boxes). Once subjects had signalled their readiness using a mouse click, each trial began with the presentation of a fixation cross for 200, 300 or 400 ms with an equal probability for each. This was followed by the central stimulus and, on the next frame, the peripheral stimulus. Following a short temporal delay, central (or peripheral) stimulus onset asynchrony (SOA), the central (or peripheral) stimulus was masked (figure 1b). SOAs were controlled such that discrimination accuracy for each stimulus in the single-task condition was held at 70% (see QUEST description below, §2e).

(ii). Dual-task condition with whole report

In the dual-task condition, both central and peripheral stimuli were task-relevant. The presentation of visual items proceeded as above; however, subjects were required to make two responses per trial, first on the central then on the peripheral stimulus (figure 1b in red). We termed this dual-task condition ‘whole-report’ to contrast it with our ‘partial-report’ procedure in experiment 2.

In order to contrast performance in the dual-task against the single-task conditions, SOAs were not updated during the dual-task condition. Instead, SOAs were fixed at the threshold duration defined by the preceding single-task block for each task type (see QUEST description, §2e). As is typical for this paradigm, we instructed subjects to prioritize performance for the central stimulus in the dual-task condition. We did not give subjects any feedback regarding their performance and did not inform them of the staircase procedure.

(iii). 8AFC response screen

Mask presentation was followed with the display of a response screen comprising eight evenly split segments (figure 1b). With a single mouse click, this screen allowed subjects to register their two-alternative-forced-choice (2AFC) discrimination response as well as a four-level subjective rating. Prior to the experiment, and during practice, subjects were verbally instructed to express their confidence from a complete guess (rating 1) to certainty (rating 4) in experiments 1 and 2 or perceptual awareness from complete invisibility (rating 1) to complete visibility (rating 4) in experiment 3. Verbal descriptors for ratings 2 or 3 were not made explicit; however, the experimenter encouraged subjects to fix these criteria across the sessions as best as they could, and use all four levels. The labels ‘sure’ and ‘not sure’ in experiments 1 and 2 (or ‘easy to see’ and ‘hard to see’ in experiment 3) were displayed at the top and bottom of the screen to remind subjects of the scale of subjective rating.

At the centre of the display, we presented the discrimination options. For the central stimuli, subjects indicated whether the target letters they had seen were all the same (S) or one was different (D) by clicking on one of the confidence segments on the side with the labels either ‘S’ or ‘D’ (figure 1b). The ‘S’ option was always on the left. For the peripheral face-gender discrimination (experiments 1 and 3), subjects indicated whether the target was either male (M) or female (F). The ‘M’ option was always on the left. For the peripheral coloured disc stimuli (experiments 2 and 3), the discrimination options were substituted with images of red–green and green–red discs (the red–green option on the left) and subjects selected a segment to indicate their percept and subjective rating.

(iv). Perceptual awareness scales

As subjective ratings for experiment 3, we employed perceptual awareness scales (PAS) [59,60], which are a more direct measure of conscious perception. For this purpose, we changed the display labels from ‘sure’ and ‘not sure’ into ‘easy to see’ and ‘hard to see’ and instructed subjects to rate ‘1’ when the stimulus was very hard to see and ‘4’ when the stimulus was very easy to see. We did not explicitly describe the ratings of ‘2’ and ‘3’, but we encouraged subjects to use all four levels when appropriate.

(v). Analysis

In addition to objective and subjective signal detection metrics to measure performance, we employed linear mixed effects (LME) modelling to examine metacognitive accuracy [57]. We also examined trade-off in performance between the central and peripheral tasks by defining a measure we call actual trigonometric altitude (TA^actual). See the electronic supplementary material for details.

(e). QUEST staircase procedure and reduced training

Previous dual-task studies employed large amounts of training to stabilize subjects' performance at the threshold SOAs [19,23–25]. For example, in Li et al. [24], training is described as usually taking ‘more than 10 h (12 000 trials of all tasks combined)’. Reddy et al. [23] trained subjects until they achieved 80% accuracy for the central letter discrimination task with an SOA below 250 ms for an entire 1 h session. This procedure took ‘between 6 and 12 h per subject’. A separate experiment by Reddy et al. [25] trained subjects for only 2 h, but all had participated in Li et al. [24] so had approximately 10 h of prior exposure to the central letter discrimination task.

To generalize conclusions of the dual-task paradigm into an untrained population, we reduced training to a minimal level by rapidly and robustly setting SOAs that yielded threshold performance levels equated between subjects. SOAs were adjusted on a trial-by-trial basis during single-task blocks using the (QUEST) adaptive staircase procedure [53]. The initial SOA for central targets was 500 and 250 ms for peripheral stimuli. We set the β parameter for QUEST to be 2 and the standard deviations to be 70% of the respective initial SOA during training. Once training was complete, the standard deviation parameter was reduced to 50 ms for both central and peripheral stimuli. On each trial of the respective single-task block, we updated either the central or peripheral SOA such that discrimination performance was fixed at 70% correct for that condition. To contrast performance in the single-task conditions against the dual-task, we did not update SOAs in the dual-task condition (figure 1c). Central SOA in the dual task was drawn from the preceding single-task central block and vice versa for the peripheral SOA.

Training in our experiments only took approximately 20 min, with subjects completing two single-task central blocks followed by two single-task peripheral blocks and 20 practice trials under the dual-task condition. This procedure reduced 12 h of training by more than 97% relative to Reddy [23].

3. Experiment 1: gender discrimination

Our first experiment examined whether gender discrimination in the near absence of attention was associated with conscious access. In addition, we employed a staircase procedure to greatly reduce the amount of training.

(a). Methods

Eight subjects (3 M, 5 F, ages 18–34) took part in experiment 1. The procedure for the experiment was identical to our general method but employed greyscale photographs of human faces as the peripheral stimulus.

A set of 65 male and 65 female faces were selected from natural crowd scenes, details of which are described elsewhere [57,61]. All were facing forward with major features (i.e. eyes, mouth, nose) clearly visible. To generate masking textures, we used 20 male and 20 female faces (out of 130 used for the experiment), each cut into 3 × 3 squares of equal size and rearranged randomly without rotation or flipping (figure 1a). Faces and masks were randomly selected on each trial.

(i). Data collection

Data collection took place over three sessions on three consecutive days. The first session consisted of training (i.e. 2 blocks x 48 trials of the single-central-letter task, 2 blocks x 48 trials of the single-peripheral-gender task and 20 trials of the dual-task), followed by two runs of the experiment. Each run comprised one block each of the single-central, single-peripheral, and the dual-task whole-report condition. The order of these blocks in each run was randomized. SOAs were updated during each single-task block and these updated SOAs were used for the dual-task block that immediately followed. Each block consisted of 48 trials. The second and third sessions skipped training and consisted of three experimental runs. This resulted in eight experimental runs overall, and thus, eight blocks for each condition (single-central, single-peripheral, dual-task) per subject.

(b). Results

(i). Face-gender discrimination in near absence of attention

Even with our minimal training procedure, we largely replicated previous findings [23,25] (figure 2a; electronic supplementary material, table S1). Objective performance (type 1 area under receiver operating characteristic curve (AUC)) for the peripheral face-gender and central letter tasks were much higher than chance (0.75 and 0.76, respectively) when they were performed simultaneously in the dual-task condition. Dual-task performance was slightly worse than each respective single-task condition (0.77 and 0.80). These differences were statistically significant (p = 0.039, for peripheral and p = 0.025, for central) according to a two-way within-subject ANOVA (attention condition (single- versus dual-task) and block as factors). Neither the interaction nor main effect of block was significant (p > 0.25; see electronic supplementary material, table S1). The TA^actual for objective performance was 0.77, indicating almost no trade-off.

Figure 2.

Conscious face-gender, but not disc-colour, discrimination is possible in the near absence of attention despite minimal training. Presented in red, the first (a–c) and second (d–f) row reflect discrimination of face-gender in experiments 1 and 3. Presented in blue, rows three (g–i) and four (j–l) reflect disc-colour discrimination in experiments 2 and 3. Experiment 3 used a blended face and disc. Each column represents a performance measure; column 1 (a,d,g,j) is objective performance (type 1 AUC) and column 3 (c,f,i,l) is metacognitive accuracy (type 2 AUC). Column 2 reflects subjective ratings (confidence judgements in experiment 1 (b) and 2 (h), PAS in experiment 3 (e and k)). Subjective ratings are plotted separately for correct and incorrect trials (lighter shades reflect mean ratings for incorrect judgements). In all panels, measures along the x-axis refer to performance on the central letter task. Measures along the y-axis refer to performance on the peripheral task. Single-task performance is plotted as upright (for peripheral tasks) and inverted (for central task) triangles near their respective axes. Dual-task performance is plotted as circles. Additional green markers are included for experiment 2 to reflect results from the partial-report procedure. Error bars signify within-subjects standard error of the mean [62,63].

(ii). Confidence remains stable in the near absence of attention

As a subjective measure, we asked subjects to rate trial-by-trial confidence on their discrimination in experiment 1 (figures 2b and 3a; electronic supplementary material, table S2). We conducted LME analysis to examine the relationship between attention and correctness on trial-by-trial confidence ratings (see the electronic supplementary material). (We also performed the analyses including block as a factor, but block was never significant in any analysis across all experiments and measures (all p > 0.25), thus we will not report it further.) For the peripheral face-gender task, the main effect of attention did not reach significance (p > 0.05 for either correct or incorrect trials), but that for correctness was highly significant (each p < 0.001 for the single and dual task). The interaction between attention and correctness was also significant, though the effect size was small (χ²(1) = 4.26, p = 0.039).

Figure 3.

Subjective ratings for peripheral stimuli as a function of attention (ST, single-task; DT, dual-task whole-report; pDT, dual-task partial-report) and correctness. Top row represents confidence ratings for experiments 1 (a) and 2 (b), respectively. Bottom row represents perceptual awareness scales (PAS) in experiment 3 (c for faces, d for discs). Lighter shading indicates mean subjective rating for incorrect responses and additional green markers in (b) for partial-report. Error bars reflect within-subjects standard error of the mean [62,63].

(iii). Metacognitive accuracy in the near absence of attention

That confidence ratings broadly correspond with subjects' accuracy implies that subjects had metacognitive insight into their decisions. We examined this relationship quantitatively by computing metacognitive accuracy (measured as type 2 AUC; see the electronic supplementary material) (figure 2c; electronic supplementary material, table S3). While metacognitive accuracies for both peripheral faces and central letters were well above chance during the dual task (0.59 and 0.60), they were significantly lower than those in the single task (0.62 and 0.64, two-way within-subject ANOVA: main effect of attention p < 0.05, ). TA^actual was 0.50, reflecting metacognitive accuracy roughly half that of single-task equivalence.

Taken together, these results confirm that despite their attention being consumed by the central letter task, subjects exhibited near intact objective accuracy and confidence ratings on their peripheral face discriminations in the dual task. While some performance decrement was seen, the magnitude was small by Cohen's conventions. Given that subjects were minimally trained in our protocol (20 min), many causes may explain this slight deficit (e.g. motor coordination errors, confusion due to task switching across blocks), and rather it is remarkable that we found highly similar results to the original findings, which required training subjects for 6–12 h. Further, no statistical analyses on performance measures showed significant effects of training (i.e. the main effect of block or interaction between block and attention), implying that our block-by-block SOA adjustment was successful in achieving stable performance. Our QUEST protocol was robust and stable enough that we saw no effects of block across any measures in experiments 2 and 3 as well (thus they are not reported further).

While subjects retain the capacity to discriminate face-gender in the near absence of attention, their metacognitive insight into these discriminations was affected in the dual-task condition. This might be due to the slight dual-task trade-off in objective accuracy and confidence ratings. We will return to this issue in our general discussion (see §6).

4. Experiment 2: disc discrimination

In experiment 2, we employed the dual-task design for coloured discs, a stimulus category regarded as indiscriminable when selective attention is diverted [22,23,64]. To test if response interference can explain this phenomenon, we used a partial-report procedure (figure 1c, see §4a).

(a). Methods

Eight new subjects (four females, ages 19–38) participated in experiment 2. Methods for this experiment were identical to our general methods apart from the inclusion of the partial-report condition and the use of coloured discs as the peripheral stimulus (figure 1a). Discs were masked by one of 10 circular, multi-coloured Mondrian patches precomputed before the experiment (figure 1a).

(i). Dual-task condition with partial report

In dual-task partial-report blocks, central and peripheral stimuli presentation proceeded as usual, but subjects were required to respond to just the central or the peripheral stimulus on a given trial. Participants did not know in advance to which stimulus they should respond, thus both tasks remained relevant while the report demands were equivalent to the single-task conditions (figure 1c).

(ii). Data collection

As in experiment 1, training and testing were conducted over the course of three sessions on three consecutive days. In session 1, training involved two blocks of 30 trials for the single-central letter task and single-peripheral disc task (120 trials in total) and 20 trials of the dual-task under whole-report conditions. No separate training was given under partial-report conditions apart from verbal and written instructions at the beginning of each block.

After training, two runs of the main experiment followed in the first session. Each run contained four blocks of 30 trials in length: one block of single-central-letter task (updated central SOA), one block of single-peripheral-disc task (updated peripheral SOA), one block of whole-report dual-task (fixed SOAs) and one block of partial-report dual-task (fixed SOAs) (figure 1c). The order of these four experimental blocks was randomized within each run. Sessions 2 and 3 consisted of three runs, resulting in a total of eight blocks of each condition per subject.

(b). Results

(i). Discs cannot be discriminated in dual-task conditions even with partial-report

In experiment 2, subjects discriminated the orientation of coloured discs in the single task, as well as the dual task under both whole- and partial-report conditions (figure 2g; electronic supplementary material, table S4). Replicating previous studies [22,23], objective accuracy (type 1 AUC) for the peripheral disc task was near chance in the dual task in both traditional whole- and our novel partial-report conditions (0.53 and 0.57, respectively) when the central letter task was prioritized (0.74 and 0.72). By contrast, both disc and letter tasks could be performed well above chance in the single-task conditions (0.77 and 0.79, respectively). Critically, we found complete trade-off (TA^actual = −0.04 and 0.02) for both the whole- and partial-report conditions (electronic supplementary material, table S4). A follow-up t-test confirmed no difference between these conditions (paired-sample t-test: t₇ = −0.608, p > 0.25, 95% CI [−0.290, 0.171]).

(ii). Confidence in discriminations for both letter and disc tasks is reduced when attention is diverted

As in experiment 1, we asked subjects to rate confidence in their discrimination judgements (figures 2h and 3b; electronic supplementary material, table S5). To examine the relationship between attention and correctness on confidence ratings in experiment 2, we conducted LME analysis (see the electronic supplementary material). For peripheral disc discriminations, the full model with interaction term differed substantially from the reduced model (χ²(2) = 68.9, p < 0.001). Subsetting data into correct and incorrect judgements revealed significant main effects of attention (χ²(2) = 769.9, p < 0.001, and χ²(2) = 146.1, p < 0.001, respectively). Subsetting by attention condition, the main effect of correctness reached significance for the single-task (χ²(1) = 111.4, p < 0.001) and dual-task partial-report (χ²(1) = 13.6, p < 0.001) conditions but not dual-task whole-report (χ²(1) = 1.3, p > 0.25). This confirmed that the relationship between correctness and confidence ratings was moderated by attention. Higher confidence corresponded with correct judgements when subjects fully attended to the disc task, but this relationship was largely extinguished when attention was diverted (figure 3b).

(iii). Metacognitive accuracy for disc discriminations reduces to chance in the near absence of attention

Metacognitive accuracy for the peripheral disc task is summarized in figure 2I and electronic supplementary material, table S6. It fell near chance-level under the dual task, in both the partial (0.53) and whole-report (0.50) conditions, though it was much above chance in the single-task condition (0.61). As we expected from our instruction to prioritize the central task, we found metacognitive accuracy of the central task was similar across conditions (approx. 0.61). In terms of trade-off analysis, we observed complete trade-off under the whole-report condition, but not in the partial-report condition (see electronic supplementary material, table S6). We will return to this and related problems regarding trade-off analysis of metacognition in our general discussion (see §6).

5. Experiment 3: blended face/disc discrimination

In our final experiment, we addressed the potential concern that gender discrimination of a face is made possible by its inherent saliency, that is, attraction of bottom–up spatial attention [38,39]. To address this issue, we developed a novel stimulus category by α-blending a face and a disc. If the saliency of faces explains our results, two aspects of the blended object, face-gender and colour orientation, should be equally discriminable under the dual task.

(a). Methods

Eight subjects participated in experiment 3 (three females, ages 21–32). Methods were identical to our general methods except that an α-blended face/disc image was used as the peripheral stimulus (see α parameter staircasing in the electronic supplementary material). The disc aspect of this stimulus was identical (but lowered in contrast through transparency) to that described in experiment 2. In order to create the face aspect, a novel set of 518 faces (half of them female) were selected from the natural crowd scenes cited above [61]. The addition of these extra faces ensured that subjects could not simply learn the face stimuli. In addition, we did not repeat a given face until every face had been presented (resulting in a maximum of four presentations of a given face across the entire experiment). As the masking stimuli, we generated approximately 3500 scrambled face textures and Mondrians, α-blended using the same technique as the blended face/disc stimulus.

(i). Data collection during the main testing runs

Training and testing for experiment 3 were conducted over three sessions on three separate days. In session 1, after the training described above, we tested subjects in two runs of the main experiment. Each run included a 30 trial block of each of our five conditions: single-central-letter, single-peripheral-disc, single-peripheral-face, dual-letter-disc and dual-letter-face tasks. We adjusted SOAs for the central letter task with QUEST in every single-central-letter block. In alternate runs of the single-peripheral blocks, we adjusted the α level (even runs) or SOA (odd runs) for the peripheral stimuli. During the dual-task blocks, we used α levels and SOAs that were updated in the preceding single-task blocks. Sessions 2 and 3 involved three runs of the main experiment each. Thus, excluding training, subjects completed eight blocks for each of the five task conditions.

(b). Results and discussion

(i). Face-aspect, but not colour-aspect, of the blended stimulus can be discriminated in the near absence of attention

When we merged a face and a disc through α-blending in experiment 3, subjects continued to possess discriminability for the face-, but not the disc-, aspect of the blended stimulus (figure 2d,j; electronic supplementary material, table S7). When subjects paid close attention to the central letter task, maintaining objective performance (type 1 AUC) at 0.81, objective discrimination performance for the face-gender aspect of the blended stimulus remained above chance (0.71) while disc-colour discrimination fell to chance (0.53). Both stimulus aspects were performed well above chance (approx. 0.75) in the respective single-task conditions. Two-way within-subject ANOVAs confirmed significant main effects of attention (single versus dual, p < 0.001), stimulus type (face versus disc, p < 0.001) and their interaction (p < 0.001) for peripheral task performance. TA^actual for face discrimination was 0.76. By contrast, TA^actual for disc-colour was 0.08, implying complete trade-off in this feature.

(ii). Perceptual awareness reduced for both features in the near absence of attention

In experiment 3, we employed a PAS (see the electronic supplementary material) to assess subjective experience [59] (figure 2e,k; figure 3c,d; electronic supplementary material, table S8). The relationship between attention and correctness on PAS ratings in experiment 3 was examined using LME analyses for each stimulus type. For blended-disc discriminations, the full model with interaction term differed substantially from the reduced model (χ²(1) = 38.73, p < 0.001). By contrast, for blended-face discriminations, a reduced model without interaction between attention and correctness did not differ from the full model (χ²(1) = 0.29, p > 0.25). The main effects of attention and correctness were highly significant when LME was applied to subset data (all p < 0.001). This confirmed that despite using the same stimuli and presentation parameters, the relationship between correctness and PAS was moderated by attention for disc-colour orientation but remained consistent between attention conditions when discriminating face-gender.

(iii). Metacognitive accuracy

The results of metacognitive accuracy based on PAS in experiment 3 were largely as expected from experiments 1 and 2 with some exceptions (figure 2f,l; electronic supplementary material, table S9). Under the dual task, type 2 AUC was high (0.59) for face-gender discrimination but near chance for disc-colour (0.53). In the single-task condition, each was individually higher than chance (approx. 0.58). Subjects maintained similar levels of metacognition for the central letter task (approx. 0.62) across all conditions. A two-way within-subject ANOVA on metacognitive accuracy for peripheral stimuli found a significant main effect of stimulus type (p = 0.032), but not for attention or their interaction (each p > 0.05). For the trade-off analysis, we identified two outlier subjects (see below). After removal of the outliers, we confirmed that while metacognitive accuracy for the face-aspect remained intact (TA^actual = 0.90 (against 0, p = 0.031; against 1, p > 0.25)), the disc-aspect dropped (0.59, p > 0.05 against both 1 and 0).

Figure 4 lists the results of objective performance, PAS and metacognitive accuracy for two subjects whose TA^actual values were greater than three standard deviations from the mean TA^actual results for n = 32 datasets pooled across experiments 1–3. Figure 4b,e shows that the two subjects did not discriminate correct from incorrect trials under the single-task condition (i.e. overlapping red (correct) and faint red (incorrect) upright triangles along y-axis), which resulted in chance metacognitive accuracy for the single-task condition (upright triangles in figure 4c,f). Because these subjects were able to discriminate between correct and incorrect trials in the dual-task condition, our method for determining trade-off resulted in TA^actual values that were massively positive (8.23 for figure 4c) and negative (−7.96 for figure 4f). We will return to this issue in our general discussion (see §6).

Figure 4.

Two outlier subjects removed for trade-off analysis of face-aspect metacognition in experiment 3. (a–c) for subject no. 17, (d–f) for subject no. 21. (a) and (d) for objective performance, (b) and (e) for PAS, (c) and (f) for metacognition. Note higher metacognitive accuracy for dual-task face discriminations than those in the single-task condition despite lower mean performance and perceptual awareness. Error bars reflect standard error of the mean over blocks. See electronic supplementary material, figure S2a–c for all subjects' data in this format. Red, correct; faint red, incorrect. Upright triangles, peripheral single-task; inverted triangles, central single-task. (Online version in colour.)

6. General discussion

In this paper, we addressed four criticisms of the dual-task paradigm: the issues of (i) metacognition, (ii) attraction of bottom–up spatial attention, or saliency, of faces, (iii) excessive training and (iv) response interference. The most important of these was metacognition: whether successful discrimination of faces in the near absence of top–down attentional amplification is achieved with conscious access. We assessed this by computing metacognitive accuracy, the correspondence between subjects' accuracy and subjective reports [54,55,57,61,65,66]. Metacognitive accuracy under the dual task has rarely been investigated (but see [56,67]). Addressing conscious access for stimuli in the dual task remains a critical limitation if this paradigm seeks to address the necessity debate [4].

We assessed the correspondence between subjects’ discrimination accuracy and subjective reports of either confidence (experiments 1 and 2) or perceptual awareness (PAS, experiment 3) using type 2 AUC [54,57,61,66,68]. For gender discrimination, objective performance, confidence ratings and metacognitive accuracy did not differ greatly between the single (i.e. attended) and dual (i.e. unattended) task conditions. When the same procedure was applied to the simple disc stimuli, performance collapsed in the near absence of attention. Any evidence of above-chance discrimination in this condition was accounted for by attentional trade-off: subjects sacrificed performance on the central task in order to respond to disc stimuli.

This pattern of results was also found when we used PAS. Subjects' PAS for gender discriminations was broadly unchanged in the dual task and corresponded well with accuracy. This suggests that phenomenology of this perceptual feature (or simply, their appearance) remains largely unchanged when attention is diverted. We note that this result and failure to differentiate PAS for disc-colour in the near absence of attention are inconsistent with a study by Rahnev et al. [69]. Using simple, grating stimuli, they found that visibility ratings for orientation judgements were lower when subjects attended.

Our findings support the claim that certain stimulus categories, such as face-gender, remain consciously visible in the near absence of attention, while other features, such as colour orientation, do not. However, accepting this conclusion hinges on our addressing the remaining criticisms of the dual task: (ii) the saliency of face stimuli, (iii) excessive training and (iv) response interference.

It is possible that subjects’ ability to distinguish gender in the periphery is a product of the inherent salience of face stimuli; a category known to attract our attention [38–40]. In experiment 3, using an α-blended, face/disc stimulus, we examined whether gender and/or colour orientation could be discriminated in the dual task. Critically, despite these stimuli being co-located and presented using equivalent SOAs in both conditions, face-gender but not colour orientation could be successfully categorized when attention was diverted to the central task. While diverting attention critically impairs subjects' ability to distinguish colour orientation, we conclude that face-gender remains discriminable not because faces attract spatial attention but because this feature remains consciously accessible in the near absence of spatial attention.

The third issue relates to training in the dual-task paradigm, which in previous studies typically required subjects to complete thousands of trials before the main experiment began [23–25]. Such extensive training may drastically alter neural circuits [70], affect consciousness [49] and impact performance [30–32]. We employed a psychometric staircase procedure in which criterion was reached in fewer than 100 trials, a reduction in excess of 97% when compared with traditional dual-task studies [23–25]. We found that face-gender discrimination was still possible in the near absence of attention, despite limited training (experiments 1 and 3).

Reducing training potentially leaves room for subjects to improve their performance during the main experiment, but we did not observe such improvement. None of our ANOVA analyses found significant main effects of block or interaction between block and other factors in any experiments or measures. Further, in experiment 3, we used more than 500 unique faces and showed each face only five times per subject over 3 days. Given these, face-gender discrimination under dual-task conditions is extremely unlikely to result from perceptual learning during the training phase, but is an inherent capacity of the visual system or perceptual learning through life.

The final issue is response-interference: could making a response on the central task result in subjects forgetting their answer for the peripheral disc task? Supporting this potential explanation, previous studies found some evidence that compared to complex stimuli, simple stimuli can be more effectively masked leaving shorter perceptual availability for subsequent reports [71–73]. Thus, in experiment 2, we examined whether disc discriminations can be rescued if the reporting procedure is simplified. This was achieved using a partial-report paradigm to reduce the load on perceptual memory. Our results clearly indicate that discrimination of colour orientation under the dual task was not rescued even when the partial-report condition minimized the influence of response interference.

(a). Limitations of the study and future directions

It is possible to think that attentional processing differs for different categories of discrimination. However, previous studies of the dual-task paradigm provide evidence that top–down attention is an undifferentiated resource [22,74,75]. For instance, Lee et al. [22] employed experiments to compare the concurrent discrimination of form, colour and motion. Interference was indistinguishable for similar (e.g. central letter task versus peripheral letter task) and dissimilar (e.g. central letter task versus peripheral motion task) task combinations, which highlights that different visual discriminations likely exhaust the same attentional capacity.

However, it is not clear if the same logic can be applied to subjective ratings and metacognition (type 2 AUC). In fact, we identified two outlier subjects (figure 4) who clearly violate the assumptions of dual-task studies. These two subjects did not distinguish correct and incorrect trials in terms of PAS in the single task but did so in the dual task. There can be several possible reasons for this behaviour.

One possibility is that the instructions for rating perceptual awareness were unclear. Because we did not include examples of invisible and visible stimuli, it is possible that subjects were not sure when to assign PAS of 1 and 4. Another possibility is that these outlier subjects used PAS ‘across’ the single- and dual-task conditions, a phenomenon we call metacognitive saturation. Inspection of these subjects' type 1 AUC and PAS reveals that their type 1 AUC and PAS for the peripheral faces were higher for the single task than for the dual task. When subjects apply a single criterion for subjective ratings across tasks of dissimilar difficulty, subjective ratings in the simple task can saturate [54,76]. This metacognitive saturation prevents type 2 measures from adequately discriminating metacognitive sensitivity in the easy task and, by comparison, inflates the metacognitive accuracy of the difficult task despite subjective ratings being lower on average.

Both of these potential issues may also relate to genuine inter-individual differences in metacognition [55,77]. In future studies, we may be able to reduce apparent individual differences by including stimuli that are clearly visible or invisible, thus setting clear reference stimuli for all participants. We can take advantage of our expedited training procedure to test many more subjects and investigate true individual differences in metacognition.

Further, these improvements permit investigation of the attentional requirements for perceiving a wide variety of stimulus features and categories. In particular, central and peripheral task discriminations could be constructed to involve stimuli that have, or do not have, supposed overlapping neural channels or receptive fields [78–80], using e.g. face stimuli in the central and peripheral tasks. This will allow one to investigate potential interference caused by overlapping neural representations of the central and peripheral stimuli, thereby clarifying whether the differentiation of these features in the near absence of spatial attention might be limited by the representational architecture of the visual system itself.

7. Conclusion

By using subjective ratings and metacognitive accuracy, we showed that certain aspects (face-gender) of peripheral vision but not others (colour-orientation) are consciously accessible in the near absence of top–down attention. This result was achieved despite minimal training and conforms with subjective reports and inferences from the dual-task literature in suggesting that the phenomenological distinction of features such as face-gender might be independent from selective attention [5,17–19,23,24,26]. Using the methods we present here, future studies might explore a range of stimulus types and features to reveal many categories of conscious perception in the near absence of attention. In doing so, we expect the distinction between top–down attention and consciousness might be clarified, permitting a deeper understanding of the functional and neuronal properties of each phenomenon [5,17,26].

Data accessibility

Data and code employed for analysing and conducting the experiments presented in this paper are available in the electronic supplementary material and on GitHub: github.com/julian-matthews.

Competing interests

We declare we have no competing interests.

Funding

L.K. and N.T. were supported by the Australian Research Council Discovery Project (DP130100194, http://www.arc.gov.au/discovery-projects). L.K. was supported by a fellowship from the Japanese Society for the Promotion of Science (JSPS P15048). N.T. was supported by Precursory Research for Embryonic Science and Technology project from the JST (3630, http://www.jst.go.jp/kisoken/presto/en/about/) and Australian Research Council Future Fellowship (FT120100619, http://www.arc.gov.au/future-fellowships).