Publications

In order to pretend that I don’t know something, I need to mentally simulate what I would do if my knowledge were different than what it is. To do this, I rely on knowledge that I have about my own mind: metacognitive knowledge. Together with Chaz and Ian, we quantified people’s metacognition by measuring their ability to pretend they don’t know something. Specifically, we had them play rounds of Hangman (“reveal the hidden word with as few letter guesses as possible”) and Battleship (“reveal the hidden ships in the grid with as few cell guesses as possible”). We then compared their behaviour in real games against their behaviour in pretend games, in which we told them what the word was, or where the ships were hidden, but asked them to behave as they would if they didn’t have this information. We found a remarkable capacity for self-simulation, but also identified important limitations: pretenders had a tendency to over-act, and were sub-optimal in their taking-in of new information.

In a commentary on a fascinating proposal by Steve and Matthias, I make a broad point about the field of consciousness sciene. I suggest that theories of consciousness are measured not only based on their scientific merits, but also based on their effectiveness as rationalization devices.

“Unconscious processing” effects are cases where behaviour is affected by events that were not consciously perceived. In the 2000s many such cases were reported in the scientific literature, but these findings were since challenged on statistical grounds. In this paper, a collaboration with Liad and Nathan, led by Itay, we point out that the statistical analyses that are commonly used to test for unconscious processing effects make an unnecessary assumption, namely that the direction of effects should be the same for all people. We developed two new statistical tests that do not make this assumption: a Sign Consistency test and an Absolute Effect Size test (the second was suggested by our Reviewer Sascha Meyen). We applied these tests to data from previous studies about unconscious perception, only to find that relaxing this assumption does not make a difference. In contrast, the same tests revealed multiple hidden effects when applied to datasets from other fields. Together, we conclude that effect of unconscious processing were not masked by individual differences.

Together with Rani and Clare, we propose a new way to model perceptual detection decisions (decisions about whether something is present or absent). Our model has only one sensor that is sensitive to the presence of a target image, and participants are assumed to observe this sensor and make decisions based on its output. This model exemplifies two principles. First, deciding that something is not there based on the sensor not activating is what philosophers call “an argument from ignorance”. And second, this makes decisions about absence strongly dependent on beliefs about the probability that the sensor would activate when the target image is present. The model fits well data from three detection experiments, and reveals that people vary in how much they take into account the properties of the sensor when making decisions about absence.

MRI machines cannot directly measure brain activity. Instead, they measures something that is closely linked to neuronal activity: blood oxygenation in different parts of the brain. To go from MRI data back to neuronal activity, brain scientists assume a model of how oxygen in the blood changes when neurons fire. But this model is sometimes off, for example in certain brain regions and among specific groups of people. Together with Roy we came up with TWISTER: a model-free approach to functional MRI experiments and analysis. We provide a proof concept and Matlab code. This paper is published as part of a special Entropy issue for Karl Friston’s 65th birthday.

In near-threshold perceptual detection tasks, participants’ task is to decide whether a faint stimulus is present or absent in a display. Current computational models describe the decision process underlying detection decisions as symmetric, as if participants report presence once the stimulus is sufficiently visible, and absence once its absence is sufficiently visible. Together with Rani and Clare, we used an occlusion manipulation to show that this description is inaccurate. While decisions about presence depend on the visibility of stimuli, decisions about absence depend not on the visibility of their absence, but on counterfactual visibility: beliefs about the visibility of stimuli that are not in fact there.

Previous research found that individuals with obsessive compulsive disorder take longer to make decisions in a visual search task, but only when they decide that the target image is absent from the display. Noam, Ruvi and I thought this was interesting because we had some ideas about possible links between obsessive compulsive disorder and being able to tell when things are not there. So we attempted to replicate this finding in an online setting. We tried twice, and failed both times. We are still not exactly sure why, but we discuss some possible explanations in the Discussion.

Consciousness is an ethically loaded term. Learning that someone is conscious changes our beliefs not only about what it is like to be them, but also about how we should and shouldn’t treat them. In this paper, we identified 3 implications of this link between consciousness and ethics. First, empirical findings in this field should inform ethics and legislation. For example, if we identify markers of consciousness in an animal, this should affect how we treat this animal. Second, because of this link, consciousness scientists are particularly susceptible to influences from ideology and societal norms. This is because we prefer theories of consciousness that align with our prior intuitions about who is conscious and who is not, but these intuitions are biased by our moral outlook and lifestyle. And finally, this link also means that the use of animal models of consciousness produces a conflict between scientific validity and ethical justification.

When participants report being aware of a stimulus, specific regions in their prefrontal and parietal cortices are activated. In this project we asked whether this brain activity was correlated with subjective visibility or alternatively with decision confidence. Together with Nadine and Steve we explored an existing dataset and found that many of the brain activations that are commonly associated with visual awareness disappear when controlling for decision confidence.

Some searches are really easy - the thing you are looking for just pops out to your attention. Interestingly, the same kind of searches also trigger an immediate perception of the absence of a target when it is not there. In this project we investigated this phenomenon of absence pop-out and found that it is independent of task experience and explicit metacognitive knowledge about search efficiency.

In detection tasks, people are slower, less confident, and less able to identify their own errors in reporting stimulus absence than presence. We thought this might be related to a more general difficulty in making inference based on the absence of evidence, a bit like interpreting p>0.05 in statistical testing. Together with Steve and Rani, we tested this idea in a Registered Report. Our findings suggest that presence/absence asymmetries in overall confidence and RT are indeed most likely to reflect a general difficulty in interpreting absence of evidence, but that asymmetries in metacognitive sensitivity are more specific to visual nothingness.

Together with my PhD supervisors Karl and Steve, we used functional MRI to ask whether brain regions that contribute to our confidence in what we see also contribute to confidence in whether we see something at all. This is the first imaging study from my PhD, and it is also the first ever pre-registered neuroimaging study to come out of the Wellcome Centre for Human Neuroimaging. We found that the same brain regions were involved in the two types of confidence, with some unexpected differences in the translation of brain activity to subjective confidence. In a second, pre-registered replication study, most of these unexpected differences were not replicated.

Working on my Master’s thesis, I got frustrated with how easy it was to get significant results out of pure noise by overfitting the analysis plan to the data. Standard pre-registration seemed like an unsatisfying solution - if we don’t trust researchers to honestly report their original analysis plans, why should we trust them that they have pre-registered their analysis plans before data collection, and not after it? Together with my brother Noam, we invented and implemented a mechanism that prevents this sort of cheating by encoding the registration plan in experimental randomization.