2026-03-25 韓国基礎科学研究院（IBS）
基礎科学研究院の研究チームは、視覚情報の「要約（gist）」処理が従来より早い段階、一次視覚野（V1）で始まることを解明した。実験ではマウスに複雑な動き刺激を分類させ、脳活動を解析。その結果、V1は単純な特徴だけでなく、動きの平均方向やばらつき（分散）といった統計情報をすでに符号化していることが判明した。これらの情報は後部頭頂皮質（PPC）に伝わり、行動判断に関わる抽象的なカテゴリー表現へと再構成される。また学習によってV1の表現も変化し、選択性の低い神経も集団として重要な役割を果たすことが示された。脳が複雑な視覚環境を効率的に理解する基本原理を明らかにし、知覚研究やAIへの応用にも寄与する重要な成果である。
Figure 1. This study shows that mouse V1 simultaneously encodes the ensemble mean and variance of motion, providing a robust summary‐statistic representation that persists despite single-neuron variability.
＜関連情報＞
- https://www.ibs.re.kr/cop/bbs/BBSMSTR_000000000738/selectBoardArticle.do?nttId=26594&pageIndex=1&searchCnd=&searchWrd=
- https://advanced.onlinelibrary.wiley.com/doi/10.1002/advs.202512369
一次視覚野および後部頭頂皮質における階層的要約統計エンコーディング Hierarchical Summary Statistics Encoding Across Primary Visual and Posterior Parietal Cortices
Young-Beom Lee, Oliver James, Gaeun Jung, Doyun Lee, Yee-Joon Kim
Advanced Science Published: 22 March 2026
DOI:https://doi.org/10.1002/advs.202512369
ABSTRACT
Despite growing evidence that the visual system pools sensory data into a summary statistical representation, the underlying neural mechanisms remain unclear. We characterized the neural coding of summary statistics at the single-cell and population levels using calcium signals imaged in primary visual cortex (V1) and posterior parietal cortex (PPC) while head-fixed mice passively viewed or classified eight mean motion directions of randomly moving dots into two categories. A small portion of neurons in both areas showed global mean motion direction selectivity beyond what would be expected from the simple summation of responses to individual dot motions. Although this selectivity was variable across stimulus variability and trials, population activity robustly encoded global mean motion direction, even though most neurons were not significantly tuned. The V1 population-level mean motion representation was dependent on stimulus variance and systematically biased toward the category center during the motion categorization task. These, along with the observed population-level neural coding of stimulus variance, suggest that multivariate V1 activity is well suited to processing summary statistics. The redundant summary statistical encodings in both V1 and PPC suggest that such information accumulates across the visual hierarchy, which may allow PPC to bind multiple levels of summary statistical representations into task-oriented category signals.
