探討主動與被動推理學習規則之神經機制

Abstract

在更新對於外在環境理解的信念時，推理是對於像人類這種智能系統非常重要的程序。信念的更新能經由被動推理，也就是透過單純觀察產生的歸納，先驗信念與後驗信念不一定有因果關係。然而從概念上來說，推理也可以是主動的，也就是經由假設驅使的行動來操縱外在環境，把信念與外在環境結果形成因果連結。於本研究中，我們檢驗了主動推理與被動推理的信念更新程序是否在神經機制上有差異，以及差異為何。實驗共有二十位參與者在功能性磁振造影（fMRI）下進行規則推理的任務，當中，每個顏色組合皆根據隱含規則被歸納至所屬類別，參與者須分別以主動及被動推理的方式習得背後隱含的規則。在主動推理的情況下，參與者能選擇顏色組合來測試他們對規則的信念是否正確。在被動推理的情況下，顏色組合則為預設。兩種推理方式在行為上的結果大致相似。然而在神經活動方面，處理顏色組合時，左角回在主動時的活動相較之下高於被動，而視覺運動區則在被動時高於主動。另外，有些腦區隨著連續答對而有更高或降低的活動，其中在主動的情況下，處理顏色組合時的丘腦、作答時的兩側角回皆升高，而回饋階段的兩側角回降低；在被動的情況下，視覺、頂葉、紋狀區域皆在處理顏色時升高，而回饋階段上枕葉區域則降低。本研究提議角回在形成假設驅使的行動過程中扮演模擬抽象景況的角色。至關重要的是，大腦中的推理可以透過主動或被動途徑進行，而這對智能系統如何理解環境具有影響。

Inferencing is a key process in intelligent systems like the human brain when updating beliefs about environmental causes and future states. Belief updating might proceed via passive inference in which prior beliefs and outcomes have no necessary causal links and posterior beliefs primarily stem from observations of associations. Conceptually, however, inference might also be active, in which hypothesis-driven actions manipulate contexts and link beliefs to environmental outcomes. In this study, we evaluated whether and how neural processing during posterior belief integration would show distinction between passive and active inferencing. 20 participants underwent a rule-inference task (RIT) functional magnetic resonance imaging (fMRI) experiment involving inferring underlying rules that map color configuration cues to target categories. In the Active condition, participants chose color cues to test their inference. In the Passive condition, color cues were predetermined. Behavioral performances were similar for both inference conditions. Nevertheless, neural responses during cue processing were higher for Active than Passive conditions in the left angular gyrus, but higher for Passive than Active conditions in visuomotor areas. Also, Active condition neural responses increased with incremental successively correct trials in the thalamus during cue processing and bilateral angular gyri during answering with feedback-related activity decreasing in the angular gyri. By contrast, Passive condition neural responses increased in visuoparietal and striatal areas during cue processing and decreased for feedback in the superior occipital areas. Our findings suggest a novel role of the angular gyri in the internal simulation of abstract contexts during generation of hypothesis-driven actions. Critically, inferencing in the brain can occur via active or passive routes which has implications for how intelligent systems represent environmental causes.