シンポジウム（Symposium）

Symposium Being Adaptive: The Role of Metacognition in Learning and Guiding Behavior シンポジウム Being Adaptive: The Role of Metacognition in Learning and Guiding Behavior 協賛：ERATO池谷脳ＡＩ融合プロジェクト

7月26日（金）8:35～9:00　第3会場（朱鷺メッセ　2F　メインホールB） 2S03m-1 What optimal and suboptimal metacognitive computations can tell us about adaptive behavior Megan Peters（Peters Megan） University of California Riverside Our brains make continuous perceptual decisions about the most likely state of the environment to help us behave adaptively and learn; typically, these decisions are accompanied by a metacognitive sense of confidence. One might expect that this sense of confidence reflects the optimal probability that the relevant decision is correct in order to encourage adaptive behavior and survival. This is certainly the dominant theory at present, and has enjoyed a sizable amount of empirical support under traditional experimental approaches and paradigms. But what if other factors, which may seem irrelevant to the task at hand, also exert influence on our sense of confidence? Recently, we have demonstrated with novel behavioral experiments and evidence from human neuroimaging that the processes underlying metacognition may diverge in puzzling ways from the assumedly optimal methods. Specifically, we have shown using computational modeling and machine learning approaches that the way perceptual information is used to compute a perceptual decision may fundamentally differ from the computations used in constructing confidence. In this talk I'll discuss several of these experiments, and how they offer converging evidence that confidence judgments take into account surprising aspects of environmental variables -- even when in some experiments they may appear optimal. Finally, I will discuss how this strategy may in fact reflect efficient encoding of environmental properties that are truly relevant for adaptive behaviors.		7月26日（金）9:00～9:25　第3会場（朱鷺メッセ　2F　メインホールB） 2S03m-2 Metacognitive control of sensory evidence accumulation Tarryn Balsdon（Balsdon Tarryn）^1,2，Valentin Wyart（Wyart Valentin）²，Pascal Mamassian（Mamassian Pascal）¹ ¹LSP, DEC, ENS, PSL University, CNRS, Paris, France ²LNC, DEC, ENS, PSL University, INSERM, Paris, France Perceptual decisions are based on the accumulation of noisy sensory evidence. These decisions are accompanied by a feeling of confidence as to whether they are correct, known as metacognition. These metacognitive signals are important for learning and the guidance of future behaviour, but here we ask whether these signals could also be important for the very process of accumulating sensory evidence for a current decision. We showed observers a series of oriented Gabor patterns, drawn from one of two circular Gaussian distributions, centred at 45 and -45 degrees from vertical. Observers were asked to decide which distribution the Gabor patterns were drawn from. In a first session the series of stimuli continued until the observer entered a response, but observers were asked to enter their response when they thought they had reached a certain target performance level. We used computational modelling to quantify how observers' behaviour deviates from ideal in this task. This analysis showed that observers' ability to control their accumulation of evidence to reach the performance targets was correlated with their metacognitive efficiency, measured via confidence ratings on decisions in a second session. We then examined how sensory evidence affects metacognitive confidence and perceptual decisions around the internal accumulation boundary of the observer. The results indicate that the accumulation of metacognitive evidence is distinguishable from the accumulation of evidence for a perceptual decision, however, these two processes of evidence accumulation interact for the efficient control of perceptual decision making.

7月26日（金）9:25～9:50　第3会場（朱鷺メッセ　2F　メインホールB） 2S03m-3 Metacognition is spared in a dual-task visual paradigm Mahiko Konishi（Konishi Mahiko）¹，Clemence Compain（Compain Clemence）¹，Jerome Sackur（Sackur Jerome）¹，Vincent de Gardelle（de Gardelle Vincent）² ¹Ecole Normale Superieure ²Paris School of Economics and CNRS When people do multiple tasks at the same time, it is often found that their performance is worse relative to when they do those same tasks in isolation. For example, error rates and response times typically increase when multitasking. People are also usually meta-aware of, and are able to accurately judge, their own task performance. This metacognitive ability has been proposed to be crucial in intra-personal cognitive control and human social interactions and has seen a surge of interest in recent years, thanks also to the recent development of bias-free measures of metacognition. However, one aspect that has received little empirical attention in comparison, is whether observers are aware of these multitasking costs. How does multitasking affect confidence and its relation to performance? We investigated this question through the use of a visual dual-task paradigm. Participants categorized both the color and the motion direction of moving dots, and then rated their confidence in both responses. Across four experiments, participants (N=85) exhibited a clear multitasking cost at the perceptual while no cost at the metacognitive level, showing instead a small benefit in the color task. Confidence in the motion task seemed unaffected by dual-tasking. We discuss several possible causes of this unexpected metacognitive benefit of multitasking and finally provide suggestions for future investigations on this matter.		7月26日（金）9:50～10:15　第3会場（朱鷺メッセ　2F　メインホールB） 2S03m-4 Metacognition simplifies reward-based learning in complex, uncertain scenarios Aurelio Cortese（Cortese Aurelio）¹，Hakwan Lau（Lau Hakwan）^2,3，Mitsuo Kawato（Kawato Mitsuo）¹ ¹ATR Computational Neuroscience Lab, Kyoto, Japan ²Dept. of Psychology, Univ of Hong Kong, Hong Kong ³Dept. of Psychology, UCLA, Los Angeles, USA Can humans learn to make rational use of their own unconscious brain representations? Previous studies have shown that reinforcement learning (RL) can operate subliminally on external masked stimuli; but the relevant information was simple and carried a single bit of information. Here we addressed a more challenging question with a novel technique based on internal multivariate representations. Subjects essentially had to learn a `hidden' brain state with many dimensions generated stochastically within the brain in order to make correct gambling choices. That is, reward contingencies were defined in real-time by fMRI multivoxel pattern analysis; optimal action policies in the gambling task thereby depended on multidimensional brain activity that is unconscious. Subjects learned to perform advantageously when their reinforcement learning processes were boosted by implicit metacognition of the hidden brain states. fMRI-based computational analyses indicate that a frontal-striatal mechanism linking basal ganglia with prefrontal cortex supports learning in tasks with high-dimensional, complex and uncertain variables. These results show that combining a psychological construct (metacognition) with computational approaches (reinforcement learning) may provide an important starting point in the development of new artificial intelligence architectures that can solve real-world problems.