Periodic Reporting for period 4 - OPTIMIZERR (Errors as cost-optimizing decisions? Redefining the origin and nature of human decision errors in light of associated neural computations)
Período documentado: 2022-11-01 hasta 2024-04-30
Making decisions, from the simplest perceptual judgment to the most complex policy-making orientation, has long been known to be prone to errors. Previous research has emphasized the ‘near-optimality’ of human decision-making. Despite its merits, this large body of work remains blind to the origin of human decision errors under uncertainty. To address this explanatory gap in the field, this project seeks to redefine decision errors not as cognitive failures, but as cognitive compromises between the accuracy of a decision and the cost associated with maintaining the precision of cognitive computations required to reach this accuracy. The first research hypothesis of this project is that human decision errors arise to a large part from the limited computational precision of inference, a cognitive limitation absent from existing theories of decision-making under uncertainty. The research objectives of this project are two-fold. First, we will test the validity of these two research hypotheses using a combination of computational modeling and multimodal functional neuroimaging of human decision-making. We will assess the degree of generality of the obtained findings by bridging research across two types of decisions historically studied independently: perceptual decisions and reward-guided decisions. Then, we will test the clinical relevance of the hypothesized trade-off between decision accuracy and computational cost in two psychiatric conditions associated with dysfunctions of decision-making under uncertainty: 1. the emergence of false beliefs in schizophrenia, and 2. the repetitive checking behavior observed in obsessive-compulsive disorder (OCD). This project, located at the interface between cognitive psychology and neuroscience, has had a significant impact on decision research but not only by revealing not whether but why we make errors in the face of uncertainty. Describing decision errors in terms of opposing cognitive pressures has started to trigger the reevaluation of the origin of psychological biases and other phenomena previously studied without considering these opposing cognitive pressures. This reevaluation may provide new targets for correcting these biases through specific procedures in the longer run. Finally, this project has also underlined the usefulness of computational modeling for understanding psychiatric disorders, by redescribing seemingly unrelated dysfunctions of decision-making under uncertainty as opposing shifts of the same cognitive trade-off.
Just at the onset of the ERC project, we had obtained findings showing that the limited computational precision of inference drives a dominant fraction of decision variability under uncertainty. These results were obtained in the context of perceptual decisions about sequences of visual stimuli. We decided to design a series of original experiments to test the generalizability of these findings for reward-guided decisions. The obtained results have underlined the degree of generality of our research hypothesis, and resulted in a key modification of the reinforcement learning (RL) models used to fit human reward-guided decisions. We have also identified the neural substrates of this learning variability in the human brain using neuroimaging data. Our results point toward an important role for the noradrenergic system in the regulation of learning precision. These results were published in in Nature Neuroscience, a broad-audience journal. I also contacted the Project Promotion team of the ERC in September 2019 and wrote with them a general-audience piece on the ERC website. Our article in Nature Neuroscience has resulted in several invitations at international research institutes to present our findings. We have also written an invited review on the origin, impact and function of computation noise for human learning and decision-making, in Current Opinion in Behavioral Sciences. We have extended this line of work by studying individual differences in learning precision across a large cohort of subjects for which non-clinical psychiatric scores are being measured. The study has recently been published in Nature Mental Health, and emphasizes the subjective and clinical relevance of the proposed accuracy-cost trade-off (our second research hypothesis). Further studies in this direction have revealed cognitive bottlenecks on human decision-making that were identified in my research proposal. In collaboration with other members of my group, using a combined behavioral-neuroimaging experiment, we have compared cue-based and outcome-based inference in healthy human participants. This research has confirmed the strong overlap between these two forms of inference at the core of perceptual and reward-guided decisions, but has also revealed selective differences between them. A first article describing the results of this study has been published in Nature Communications, a broad-audience journal, and a second one in eLife. Last, in collaboration with medical collaborators, I have conducted research to validate the clinical relevance of the experimental approaches developed in my research project. In particular, we have applied a new cognitive task to severe OCD patients and observed selective differences in the outcome-based condition of the task.
The conducted research has provided support for the core hypothesis of my proposal: that a large fraction of the variability of human decisions under uncertainty is due to the precision of computations during inference, not from action selection. This key finding is particularly important given the breadth of the literature which assumes (erroneously given our recent findings) that human decision variability arises from stochastic action selection policies. I have also tested the subsequent hypotheses of my proposal by testing whether this limited precision of cognitive computations results in specific cognitive strategies for decision-making under uncertainty. In this respect, we have tested whether human participants seek to ‘compress’ sensory information onto decision-relevant evidence during inference - something which has not been tested to date. The findings confirm the hypotheses and go beyond the state-of-the-art by showing that the strategies used by humans to make decisions under uncertainty are shaped by the limited precision of cognitive computations uncovered in the first period of my project. We have also shown, as predicted in the proposal, that humans optimize a trade-off between the target accuracy of an uncertain decision and the precision of cognitive computations used to achieve this accuracy. The results have been published in Science Advances, a broad audience journal, in 2023. Regarding clinical aspects of my project, we have shown that OCD patients are altered in their decisions under uncertainty only when this uncertainty concerns the outcomes of their actions. These findings provide novel evidence for a specific impairment of cognition in severe OCD. Indeed, existing research has not been able to make such specific observations due to the lack of a cognitive task. We have also identified alterations of inference in positive symptoms of schizophrenia using a pharmacological model of early-stage psychosis tested in healthy volunteers. These results have been published in Nature Communications, a broad audience journal.