<ハイブリッド開催に変更>94th CiNet Monthly Seminar: Hakwan Lau "Comparing Representational Geometries Across Tasks, Brain Regions, and AI Systems", Aurelio Cortese "A metacognitive architecture for credit assignment in the human brain"

CiNet Monthly Seminar

2026年12月11日(金)
13:00-15:00
CiNet棟大会議室およびオンラインのハイブリッド形式にて開催いたします。

会場で参加される方は、直接CiNet棟1Fの大会議室にお越しください。
オンライン参加をご希望の方は こちらのフォーム からお申し込みください。
【申込締切:12/10(木)正午】
※所属が未記入や略称での記入により確認できない場合は、ご参加いただけない可能性があります。
※当日のアクセス情報などは、12/10中にメールにてお知らせします。

ーーーーーーーーーー

演題:Comparing Representational Geometries Across Tasks, Brain Regions, and AI Systems

韓国
Institute for Basic Science, South Korea
Center Director
Hakwan Lau

Abstract:
Whether two visual objects appear similar depends on the feature dimension under consideration. An apple may seem more similar to an orange than to a fire engine because both are fruits, whereas a comparison based on colour could yield a different pattern. For this reason, similarity judgments are often regarded as ill-posed unless the relevant dimension is specified. We show, however, that there is a principled and natural sense in which the similarity of two stimuli, in terms of subjective appearance, can be defined: by an observer’s own psychophysical sensitivity to discriminating them. Building on this finding, we introduce methods for recovering sensitivity-based representational geometries from neural data, capitalizing on both the phenomenon of repetition suppression and the method of multivariate pattern analysis. We show that such representations can be identified across brain regions at multiple levels of the visual-processing hierarchy. Comparing these geometries across tasks, cortical regions, and artificial-intelligence systems may provide a more precise way to assess whether their underlying computations are aligned.

ーーーーーーーーーー

演題:A metacognitive architecture for credit assignment in the human brain

韓国
Institute for Basic Science, South Korea
Associate Professor
Aurelio Cortese

Abstract:
In 1846 Le Verrier saw Uranus depart from its Newtonian orbit. He assigned the error to the model, not the telescope, and found Neptune. Every agent with a world model faces this problem: an error is useful only if assigned to the right level. I will present work on how the brain uses metacognitive insight into each level, the state of the world and the percept, to decide whether the model guiding action was wrong or the evidence misread. In a dual-confidence task participants tracked uncertainty over both, and after an error the two confidences decided, in opposite directions, whether to revise the inferred state (or stay with the current strategy). We developed a noisy Bayesian observer model that apportions each error by the relative reliability of the two levels, further showing that participants with lower metacognitive ability also display noisier inference and error allocation processes. MEG, fMRI, and intracranial data show the brain divides the task space hierarchically and performs this arbitration via a prefrontal-congulate circuit, with computations at specific frequencies. Metacognition, I will argue, is how humans audit their world model and learn from failure.

担当 : Rob Mok