| 摘要: | 「統計學習」指的是生物在自然情況下、通過適應環境中的規律性而呈展出的學習效果。儘管這一概念涵蓋了廣泛的學習現象，有研究者提出這些現象皆可用一系列記憶歷程來描述，主要涉及資訊分塊和整合的過程。現有的行為與功能性神經造影研究證據皆表明，統計學習並非由單一的神經機制所支持，而是涉及幾乎所有參與當下信息處理的神經網絡。由於海馬迴所處的信息層級，以及適合進行快速且複雜運算的神經活動特性與結構，在學習與記憶的動態過程中發揮著獨特且重要的作用。當前的研究者普遍認為，海馬迴透過發展共享的記憶痕跡（memory engram）細胞來促進關聯（associations）的習得，而它的模式完成（pattern completion）功能能促進並強化關聯學習；同時，它的模式分離（pattern separation）功能主要用於維持資訊的精確和獨特性，與關聯學習互補。然而，我們認為模式分離功能可能也在關聯習得的過程中產生貢獻。為探究這項可能性，本研究一共採用了兩種取向的實驗：一是檢視個別差異以探討統計學習與其他能力（特別是海馬迴的模式分離功能）之間的相關性，二是觀察與統計學習過程相關的大腦活動變化，特別聚焦於支持模式分離功能的海馬迴和支持基礎視覺處理歷程的主要視覺皮質區。
實驗一旨在了解個體在統計學習作業中的表現是否與受到海馬迴支持的模式分離功能有關。我們採用觀察個體差異的研究方法，透過受試者在一系列作業中表現的相關性來探究兩種認知功能在機制上是否存在顯著的相關性。實驗一分成1a與1b兩個部分，分別以105名年輕受試者和47名長者作為研究對象，來進一步了解老化潛在可能造成的機制變遷。受試者進行的實驗作業包含有：以視覺或聽覺呈現規律的兩項經典統計學習作業、一項動作序列學習作業、一項對模式分離功能敏感的記憶測驗與三項傳統記憶測驗、兩項工作記憶測驗，以及兩項非語文流體智力測驗。
實驗1a的結果顯示，受試者在兩種不同感官輸入下的統計學習作業表現皆與他們在記憶測驗中的相似物區辨（lure discrimination）能力指標呈顯著相關。這項指標被認為與海馬迴的模式分離功能相關，這意味著個人海馬迴在替個別經驗建立正交化編碼的性能優劣或傾向強度，可能會直接或間接地影響他們進行統計學習的效益。我們進一步從實驗1b探究老化對統計學習機制造成的影響中觀察到，年長受試者的相似物區辨能力仍舊與其在視覺統計學習作業中的表現相關，但與其在聽覺統計學習作業中的表現則失去相關性，反而變得與外顯配對關聯學習的能力相關。另外，在年長受試者身上，我們原先在年輕受試者身上觀察到的視覺和聽覺統計學習作業表現之間的相關性也不復存在。這可能表示，長者在進行聽覺統計學習時的表現更受其聽知覺能力的影響，導致由海馬迴機制所支持的模式分離功能對整體表現的影響降低，因而無法觀察到聽覺、視覺統計學習作業共享相同的領域共通（modality-general）學習機制。
實驗二有兩個目的：首先，我們認為海馬迴的模式分離功能有助於減緩偶然習得的配對關聯造成的混淆，想對此進行驗證。其次，我們試圖探索統計學習過程中不同腦區隨時間展現的活動消長。為此，我們進行fMRI實驗，招募了37名受試者，讓他們觀看一系列看似隨機呈現、實際上是以每三項為一組有固定時序關係的「三連體」為隱藏結構呈現的幾何圖形，並記錄他們在熟悉這些規律、也就是視覺統計學習的期間與前後的大腦活動。一方面，我們透過表徵相似性分析檢查在熟悉「三連體」規律結構後，在兩個三連體邊界處相遇的項目是否會因應區隔不同時序組合的需求，表現出表徵「差異性」的強化。另一方面，我們直接觀察海馬迴、三個內側顳葉皮質、以及視覺處理皮質（合併的V1和V2、餘下的lateral occipital cortex、fusiform gyrus、inferior temporal gyrus）在學習歷程中的活動。由於位於三連體內不同時序位置的圖形自然帶有不同屬性：處在「第一」和「第二」個項目位置的圖形具有預測力、而在「第二」和「第三」個項目位置的圖形是可被預測的，因此我們意圖探索這些位於視覺信息處理系統中不同層次的腦區，對於「有預測力」和「可被預測」的刺激項目在學習歷程中是否、又會如何展現不同的活動與互動趨勢。
在實驗二的結果中，我們沒有觀察到任何腦區根據三連體的分組展現顯著的表徵相似性變化。值得注意的是，儘管行為測驗的結果顯示受試者確實有學習到某些刺激材料之間的序列規則，我們也未能重現之前腦造影研究中的結果，即對於在統計學習作業中屬於同一個三連體的刺激材料，其海馬迴系統內表徵的相似性會提升。考慮到與之前的研究相比，我們使用了更簡單、容易記憶的視覺刺激，因此可推測不同類型刺激材料的表徵可能具有不同程度的可塑性。關於第二個目的，我們發現在統計學習期間，不同腦區的活動動態存在差異。不同層級的視覺處理皮質在整個學習過程都有展現顯著活動，其中較高階的視覺皮質（除了合併的V1和V2之外）的活動強度有隨時間下降的趨勢。海馬迴在大部分的時間沒有展現顯著一致的活動，但在學習過程中期對於「有預測力」的刺激項目的活動強度有所提升。這意味它可能在使用模式完成（pattern completion）功能產生預測信號，透過預測編碼來促進統計學習的發生。內側顳葉皮質隨時間的變化趨勢在整體上與海馬迴更接近，但在學習過程中期也對「可被預測」的刺激項目提升活動強度，並且也對「可被預測」的刺激項目展現隨時間的活動程度下降。這複雜的活動模式意味這些腦區可能參與在多種不同機制中，有待後續研究釐清。
總結來說，這項研究的結果，無論是來自行為或是腦造影的實驗證據，皆支持海馬迴在統計學習歷程中有發揮作用，同時這些結果進一步地顯示海馬迴的「模式分離」和「模式完成」兩種功能可能都有助於更好地掌握環境中的規律性。當前的研究者通常將海馬迴參與在統計學習過程中發揮的效用與其模式完成功能連結，後續的研究值得深入探究海馬迴是否在學習的不同階段發揮著不同作用。;“Statistical learning” refers to the learning effects exhibited by organisms through adaptation to regularities embedded in their environment under natural settings. While it encompasses a broad range of learning phenomena, researchers have proposed that they can generally be categorized as the extraction and integration of information, with memory processes serving as foundational building blocks for their emergence. Accumulating evidence from behavioral and functional neuroimaging studies suggests that SL is not supported by a single mechanism, but involves nearly all neural networks engaged in the processing of task-related information. The hippocampus, by virtue of its hierarchical position as well as its neural characteristics and architecture, plays a unique and significant role in the dynamic processes of learning and memory. Current researchers generally agree that the hippocampus facilitates the acquisition of associations by developing shared memory engram cells, with its pattern completion function having a synergistic effect. Conversely, its pattern separation function is considered to serve a complementary purpose, primarily responsible for maintaining the precision and distinctiveness of information. However, we speculated that pattern separation may also contribute to the acquisition of associations. To explore this possibility, the current study employed two experimental approaches: First, examining individual differences to investigate correlations between SL and other individual abilities, particularly the pattern separation function of the hippocampus. Second, exploring changes in brain activity associated with SL processes, focusing on the hippocampus (which supports pattern separation) and visual cortices (which are essential for visual processing).
Our first experiment sought to determine whether variation in SL performance across individuals can be associated with the efficacy of their hippocampal pattern separation function. We adopted an individual differences approach to examine whether a significant mechanistic link exists between these two cognitive processes, by analyzing the relationships between participants’ performances across a series of tasks. Experiment 1 comprised two parts, 1a and 1b, involving 105 young adults and 47 older adults, respectively, to further understand the potential the impact of aging on SL mechanisms. Participants performed tasks including two conventional SL tasks in visual or auditory modalities, one motor sequence learning task, a memory test sensitive to pattern separation, three traditional long-term memory tests, two working memory tests, and two fluid intelligence tests.
Results from Experiment 1a demonstrated an interconnection between SL abilities across sensory modalities, with performances on both SL tasks correlated with lure discrimination performance, which is closely related to the competence of hippocampal pattern separation function. This suggests that the superiority of the hippocampus to encode distinct experiences uniquely may influence the efficiency in SL. Experiment 1b further revealed that, among older adults, while the performance in visual SL remained correlated with the ability to discriminate lures in memory, those of auditory SL didn’t and, instead, showed an increased connection with the ability to explicitly learn (auditorily presented) paired associations. Furthermore, with aging, SL abilities across sensory modalities become less predictive of each other. These results indicate that, in elderlies, the capability to perceive and process auditory information may become a dominant factor influencing the variability of SL performance. Consequently, the impact of hippocampal pattern separation appears reduced, leading to a lack of observed modality-general SL mechanisms.
Our second, fMRI experiment had a dual purpose: Firstly, we aimed to identify a signature indicating that hippocampal pattern separation takes a part in SL processes. Secondly, we sought to explore the temporal dynamics across different brain regions during SL. To these ends, we recruited 37 participants, presented them with a series of geometrical shapes that were implicitly structured as triplets (i.e., three items with a fixed temporal order). On one hand, we examined whether representational patterns of shape stimuli sitting astride triplet boundaries would undergo differentiation after being familiarized with such regularity structure, implying pattern separation aids in reducing confusion caused by accidentally experienced associations. On the other hand, we explored whether and how regions at different levels of the visual information processing hierarchy, including the hippocampus, three MTL cortices, and four visual cortices (combined V1 and V2, remaining lateral occipital cortex, fusiform gyrus, inferior temporal gyrus), would exhibit different patterns of activity changes throughout the familiarization phase, possibly contingent upon the predictiveness (the first and the second items within a triplet can predict upcoming stimuli) and predictability (the second and the third items within a triplet can be predicted) of stimuli.
The results of Experiment 2 did not detect any significant changes in representational similarity across brain regions after learning. Notably, despite behavioral evidence of learning triplet structure (or sequential associations among stimuli), we were unable to reproduce previous results showing increased representational similarity for associated stimuli (i.e., items belonging to the same triplet) within the hippocampal system. Considering the use of simpler and more memorable visual stimuli compared to previous studies, we speculate that the plasticity of internal representations may vary depending on the properties of the stimulus materials. Regarding the second purpose, we found that activity dynamics during the learning phase vary across brain regions. Visual cortices were consistently engaged throughout the SL process, with its higher-level parts (excluding combined V1 and V2) showing a decline in activity over time, regardless of whether the stimuli is predictable. The hippocampus showed no consistent participation for most of the time, but increased its activity for predictive stimuli in the middle stages of the learning phase, possibly engaged in predictive coding process, facilitating perceptual processing by generating prediction signals through pattern completion. The MTL cortices exhibited complex activity patterns, roughly resembling those of the hippocampus, but also increased activity for predictable stimuli during the middle phase and decreased its response intensity over time, suggesting their involvement in multiple mechanisms that require further investigation.
In conclusion, the results of this study, both from behavioral and neuroimaging evidence, indicate that the hippocampus plays a role in SL processes. Furthermore, our results suggest that both the “pattern separation” and “pattern completion” functions of the hippocampus contribute to the better grasping of regularities in the environment. While current researchers predominantly associate the hippocampus’s involvement in SL with its pattern completion function, future studies should consider exploring the potential for the hippocampus to serve various functions at different stages of the SL process. |
