**4. Discussion**

A group of volunteers was tasked to classify Necker cubes of different ambiguity within 40 min. The subjects reported the orientation (left or right) of each presented cube, while the stimulus morphology ranged from low ambiguity (LA) to high ambiguity (HA). By observing behavioral responses, we found that decision time decreased with the time on task for both HA and LA stimuli (Figure 2D). At the same time, the subjects improved the correctness of the interpretation of the HA stimuli, but not the LA stimuli (Figure 2F). Analysis of the EEG spectral power on the sensor level and in the source space revealed an increase in the prestimulus power at 9–11 Hz with the time on the task (Figures 3B and 4B). Finally, we found that the prestimulus EEG power negatively correlated with the decision time to LA stimuli (Figure 5A) and the number of erroneous responses to HA (Figure 5D) stimuli.

First, we hypothesized that the prestimulus EEG power reflects changes in a person's condition. The condition, in turn, affected the performance of processing the ongoing visual stimulus [11,16]. Thus, our results showed that the high pre-stimulus 9–11 Hz EEG power predicted faster decision times and greater accuracy. It is worth noting that we compared EEG power and behavioral estimates between time segments, each of which lasted 10 min. Therefore, we have associated the described effects with slow changes in the observer's state. Taken together, we proposed a possible application of our findings in passive brain-computer interfaces to monitor the human's condition and predict decision speed and errors.

Along with possible practical applications, our result can help to reveal specific features of ambiguous perception. To start a discussion on this aspect, we will look at the limitations of our experimental design. It consists of two confounding variables, mental fatigue, and learning; both can be time-dependent. Thus, the revealed EEG changes practically do not reflect one of these factors. At the same time, we suppose that fatigue

and learning affect the observer state regardless of the stimulus ambiguity. In contrast, our results show a different time-on-task effect on processing HA and LA stimuli. For instance, we revealed a reduction in error rate for HA stimuli only. In addition, the prestimulus EEG power negatively correlates with the decision time to LA stimuli, not to HA stimuli, and negatively correlates with the error rate for HA, but not for LA stimuli. Another limitation is the small sample size. Thus, there is a risk that the individual characteristics of the participants, e.g., gender and age, will affect their perception of ambiguous stimuli. For our group, we observed no gender and age effects on the decision time and the ERSP, due to the almost uniform distribution of these factors between participants. At the same time, we expect that another group of younger or older subjects may demonstrate different effects on the behavioral and neural activity levels.

The decision time (DT) may decrease due to neural adaptation (NA), which occurs when the same visual stimulus is repeatedly presented within a short interval and causes a decrease in neural response to repetitive versus non-repeated stimulus [33]. The NA is thought to arise from at least two types of neural activity. One explanation is that only the part belonging to the neuronal ensemble is sensitive to stimulus recognition. Thus, the neurons that are not critical for stimulus recognition decrease their responses when the stimulus reappears, while on the contrary, neuronal populations carrying essential information continue to give a robust response. As a result, the mean firing rate decreases due to stimulus repetition [34]. An alternative explanation is that stimulus repetition reduces response in the time domain [33]. According to this theory, a neural network that processes sensory information responds faster to a repetitive stimulus than to a new stimulus, i.e., a stable response. The network connections involved in the response creation were reinforced by the previous presentation of the same stimulus [35]. The NA affects the neuronal response in the occipital [36], parietal [37], and frontal [38] cortex areas in single-unit data, and on the sensory level. As known from the literature, NA, as a rule, reduces the stimulus-related EEG/ECoG response to stimuli. Here, we did not consider the post-stimulus EEG power and did not report such signs of NA. At the same time, an increase in the prestimulus EEG power may reflect the preactivation of sensory neurons. We suppose that in this preparatory state the neural ensemble exhibits less activation in response to the stimulus. Further research should verify this hypothesis by examining the post-stimulus activity as a function of the time-on-task.

Another potential explanation is the vital role of alpha-band oscillations for visual perception. Our results showed that an increase in the 9–11-Hz band power correlates with enhancing processing performance. In contrast, many studies have reported negative effects of alpha power on processing performance. For example, the authors of Ref. [12] reported that moderate to low alpha signal strength in the prestimulus period is beneficial for sensory perception. In contrast, a recent review [39] emphasizes that high alpha power facilitates perception by suppressing irrelevant input and generating predictors in the visual cortex. The latter was confirmed by observing an increase in the prestimulus alpha power when participants could predict the identity of the forthcoming stimulus [40].

Finally, the role of alpha-band oscillations depends largely on their incident brain region. For instance, the authors of Ref. [41] provided evidence that right temporal alpha oscillations play a crucial role in inhibiting habitual thinking modes, thereby developing creative cognition. Another work showed that observing a Necker cube can improve subsequent creative problem-solving [42]. In line with these works, we supposed that increasing 9–11 Hz power in the right temporal region reflects a developing ability to inhibit obvious associations. According to Ref. [42], the latter may be a biomarker of neural processes facilitating creative problem-solving.

We hypothesize that NA affects the bottom-up processing, while the other two represent the top-down processing components [43,44]. Assuming that NA facilitates the bottom-up processing, we provide a possible explanation for the negative correlation between the prestimulus EEG power and the decision time to LA stimuli. The morphology of inner edges unambiguously determines the orientation of the LA stimulus. Consequently, during the LA stimulus processing, the bottom-up component prevails [45]. If so, neurons in sensory areas receive the information needed to make the right decision. Repeated stimulation pre-activates these sensory neurons, which leads to a decrease in decision time to LA stimuli. In contrast, the HA stimuli processing may require top-down mechanisms, because the morphology remains similar for the cube being either left or right-oriented [20]. To explain the lack of correlation between the prestimulus EEG power and decision time to HA stimuli, we also assume that during HA processing, the top-down component predominates for most of the time interval, while the bottom-up one may be limited to an earlier and shorter time window [16]. Thus, by facilitating the bottom-up processing, NA has little or no effect on the overall decision time for HA stimuli.

Summing up, we can say that the increasing prestimulus EEG power can reflect NA in sensory neural networks encoding the Necker cube morphology, which explains the negative correlation between the EEG power and the decision time to LA stimuli. At the same time, decision time to HA stimuli also decreases with time on task, but barely correlates with the prestimulus EEG power. This behavioral effect is probably the result of neural processes acting after the stimulus onset and relies on the integrative dynamics, rather than the EEG power modulation in a particular area. In our future studies, we will address this issue by considering the functional connectivity evolution during the experiment.

By examining the error rate, we found that the observer responded more correctly to LA stimuli. The number of erroneous responses is less than 2%, and remains unchanged during the experiment. In contrast, the number of incorrect responses to HA stimuli decreased with time on task and negatively correlated with the prestimulus 9–11 Hz EEG power. This effect may be a result of the top-down mechanisms, e.g., predicting the identity of the forthcoming stimulus or creative thinking. As discussed above, these processes also accompany increasing alpha-band power and relate to ambiguous stimuli processing.

## **5. Conclusions**

During the Necker cube classification, participants decrease their decision time with the time on task, regardless of the stimulus ambiguity. At the same time, they improve the correctness of interpretation only for the highly ambiguous stimuli. EEG analysis has revealed growing prestimulus alpha-band power with the time on task. We have found that the prestimulus EEG power negatively correlates with the decision time for the stimuli, with low ambiguity and the number of erroneous responses to highly ambiguous stimuli.

We suppose that repetitive stimuli presentation affects top-down and bottom-up processing mechanisms. Thus, it may cause the neuronal adaptation of the sensory neurons facilitating bottom-up processing. For the low ambiguity, the bottom-up component dominates; therefore, decision time correlates with the prestimulus EEG power. Increasing alpha-band EEG power may also reflect modulation of top-down components, e.g., the ability to suppress irrelevant information and form the predictors of ongoing stimulus. For the high ambiguity, the top-down processes dominate; therefore, the correctness rate correlates with the prestimulus EEG power.

Finally, our results may add to uncovering the associations between the current human condition and their performance in solving ongoing tasks. It is essential for BCIs to aim not only at the detection, but also the prediction of the human states. These BCIs will give rise to the artificial intelligence systems that assist or alarm when detecting a high probability of critical errors.

**Author Contributions:** Conceptualization, A.N.P.; Data curation, A.E.H.; Formal analysis, S.A.K. and A.N.P.; Funding acquisition, A.E.H.; Investigation, A.K.K. and S.A.K.; Methodology, V.A.M. and A.E.H.; Project administration, A.E.H.; Software, A.K.K.; Validation, V.A.M.; Visualization, S.A.K. and V.A.M.; Writing—original draft, V.A.M.; Writing—review and editing, A.N.P. and A.E.H. All authors have read and agreed to the published version of the manuscript.

**Funding:** This research was funded by the Russian Science Foundation (Project 19-72-10121) V.A.M. supported by the Russian Foundation for Basic Research (19-32-60042) in part of the experimental paradigm development and behavioral data analysis. S.A.K. is supported by the President Program (MD-1921.2020.9) in the part of source reconstruction.

**Institutional Review Board Statement:** The study was conducted according to the guidelines of the Declaration of Helsinki, and approved by the Institutional research ethics committee of the Innopolis University (protocol 2 from 18 January 2019).

**Informed Consent Statement:** Informed consent was obtained from all subjects involved in the study.

**Acknowledgments:** The authors graciously acknowledge the subjects for their contribution to this research.

**Conflicts of Interest:** The authors declare no conflict of interest. The funders had no role in the design of the study; in the collection, analyses, or interpretation of data; in the writing of the manuscript, or in the decision to publish the results.
