This doctoral dissertation presentation introduces a study that explores metacognition, intended as confidence in decision making, by employing a combination of psychophysics, functional magnetic brain imaging (fMRI), machine learning, and mathematical modeling.
The first part of the presentation focuses on the representation problem and the two main models on the computation of confidence by the brain. Specifically, in this study selected brain regions were examined with fMRI and multivoxel pattern analyses, on the basis of their importance for visual information processing as well as decision confidence. There is converging evidence in the literature that frontoparietal areas may be crucial for confidence judgements, but also that confidence and the content of perception may evolve directly through the same neuronal circuitry. This has led to an open debate, as yet unresolved. Results presented here indicate a complex scenario, and warrant a more refined experimental approach through the use of decoded neurofeedback (DecNef).The second part of the presentation discusses about the online manipulation of multivoxel representations of confidence in selected frontoparietal regions of the brain. Results indicate that confidence was changed bidirectionally, and critically, without leak of effects on task accuracy. Further psychophysical analyses excluded an effect driven by response criterion. Importantly, successful induction of multivoxel activation patterns correlated with the changes in confidence ratings, providing a causal link between neural substrates and behavioral outcomes. These results provide evidence for the hypothesis that confidence emerges as a late-stage read-out.
The third and last part of the presentation examines a further, more general aspect of any online neurofeedback manipulation in within-participants designs. Because each participants underwent neurofeedback training twice, each time with opposing signs, an important aspect pertains to the elucidation of its dynamics. Nonlinear mathematical modeling was employed to formally demonstrate bidirectional effects of DecNef, as well as learning interference between sessions. Furthermore, these findings are interpreted in the context of reinforcement learning, and they provide important constraints on real-time multivariate fMRI applications to basic neuroscience as well as therapies for cognitive and mental disorders.