人们依靠多种不同的参照系（Frame of reference，FOR）来表征和更新不同物体的空间关系。根据心理语言学研究，参照系可以分为自我参照系（Egocentric FOR，EFOR）、内在参照系（Intrinsic FOR，IFOR）和环境参照系（Allocentric FOR，AFOR）。这些不同的参照系共同影响着人类的空间认知，但目前关于人类如何表征和加工多重空间参照系这个问题还不清楚，如多重参照系是以并行方式还是以序列方式加工。本文以认知框架理论为指导，试图从宏观的表征水平和微观的神经计算水平分别阐明多重参照系的加工机制。在表征水平，我们通过提出空间信息加工机制可划分为注意分配和反应选择两个过程模块，结合行为、眼动、脑电等技术手段，寻找不同参照系在两个过程模块并行表征与加工的行为和神经证据，以验证多重空间参照系的并行加工机制。在神经计算水平，结合计算神经建模，本文提出神经元群组通过分布式表征实现并行加工不同参照系，并进一步推出在表征水平上对多重空间参照系注意分配和反应选择的两个加工过程源于神经计算水平上神经元群组的预测学习，即通过对输入信息进行时空整合以使得神经元群组对输入信息的输出与预测的差异达到最小化的过程。
本研究主要采用双大炮射击目标实验范式，即首先呈现红色和蓝色的两个大炮和八个潜在目标点，随后一个点闪烁黄色光圈成为目标点，要求被试通过左右按键判断与目标点颜色相同的大炮指向目标点的最短旋转方式（左键-逆时针旋转，右键-顺时针旋转）。通过潜在目标点颜色比例变量调控不同内在参照系（两个大炮）的优势程度（线索效应）来考察不同参照系的注意分配过程，通过两个大炮夹角变量和目标大炮朝向变量分别调控两个内在参照系之间（IFOR-IFOR）的反应冲突及自我参照系与内在参照系之间（EFOR-IFOR）的反应冲突（冲突效应）来考察不同参照系的反应选择过程。
我们实验发现：一、在注意分配过程上，个体会快速根据背景信息及其变化而分配注意到不同参照系上并影响后期的反应选择。具体通过线索效应来体现，即高概率目标点相比于低概率目标点，个体表现出更短的反应时、更低的错误率、更正的中央区P3（396-726 ms）波幅，在晚期的400-800 ms时间段上alpha（8-13Hz）和beta（13-20Hz）频段更低的频谱波动（ERSP）和theta（4- 8Hz）、alpha和beta频段更低的试次间相关一致性（ITC）。在时间分布上，在线索呈现阶段，相比于非优势参照系，个体表现出对优势参照系更长的注视时长和更高的注视比率；当目标点出现后，与目标点颜色相同的参照系优势程度提升到最高，从而获得最高的注视比例，与它在线索阶段的优势程度无关。二、在反应选择过程上，我们观察到IFOR-IFOR及EFOR-IFOR的反应冲突，并且两种冲突间存在交互作用（冲突条件相比于非冲突条件反应时更慢，错误率更高。交互作用表现为EFOR-IFOR冲突效应量在IFOR-IFOR冲突条件下相比于非冲突条件下显著增大）。脑电结果也验证了这一效应，并进一步发现两种冲突表现出了独特但又共享的神经活动模式。脑电结果发现两种冲突的冲突条件相比于非冲突条件，N2（276-326 ms）和P3（396-726 ms）都有更低的波幅，并且在P3后期（561-726 ms）波幅上两种冲突存在交互作用。时频结果发现在早期的200-400 ms时间窗口，IFOR-IFOR冲突独立调控beta频段，EFOR-IFOR冲突独立调控theta频段。在晚期的400-800 ms时间窗口，两种冲突共同调控alpha和beta频段。
计算神经建模发现，在训练成功后，模型在计算步长上表现出与反应时相似的线索效应和冲突效应。进一步分析发现，神经元群组能够通过预测学习从输入信息中抽取出不同的参照系表征（不同参照系的敏感神经元呈现出不同的离散分布式排列）。其中高优势参照系敏感神经元激活更高可能导致了线索效应的产生，而不同参照系表征彼此共享的一部分敏感神经元使得不同表征间发生混淆可能是产生冲突效应的原因。时间分布结果表现出与眼动注视比例相似的结果，即不同参照系表征的敏感神经元激活水平随着它们优势程度的变化而变化。
本研究为探索人类如何加工多重空间参照系提供了表征水平和计算神经水平的解释，实验结果支持了多重参照系并行加工机制的假设，并进一步试图在两种水平间建立连接，即计算神经水平的神经元间时空整合的预测学习可能是形成表征水平的注意分配和反应选择过程的底层基石。

People adopt multiple frames of reference (FOR) to represent and update the relationship of objects in the complex environment. Based on the psycholinguistic research, FORs can be classified as egocentric FOR (EFOR), intrinsic FOR (IFOR), and allocentric FOR (AFOR). Human spatial performance is determined by the interaction of all relevant FOR-based representations. However, how human represent and process these FORs remain elusive, such whether the process among FORs is in a serial way or in a parallel way. Here, we addressed this issue at two different levels. At the macroscopical representative level, we investigated the parallel processing mechanisms underlying the attention allocation and response selection among multiple FORs utilizing behavioral, eye-tracking, electroencephalogram (EEG). While at the microscopical neural computational level, we adopted the computational neural network technique to show how neural population distributively represented and parallelly processed the FORs. We hypothesized that at the neural computational level, the predictive learning, a process of spatiotemporal integration that minimizes the discrepancy of the expectation and the outcome at each time point, might primarily drive the emergence of the processes of attention allocation and response selection at the representative level.
We adopted a modified two-cannon paradigm. At the beginning of a trial, eight colored dots surrounding two colored cannons appeared. Then one dot became the target as indicated by a flashing yellow ring. Participants were asked to rotate the target cannon (the cannon of the same color as the target) with a smaller angle, either counterclockwise (pressing the left key with the left index finger) or clockwise (pressing the right key with the right index finger), in order to shot at (point to) the target as quickly and accurately as possible. We manipulated the color proportions of the surrounding dots to regulate the salience of FORs (cueing effect) to investigate the attention allocation. We manipulated the two-cannon angle, and the orientation of the target cannon to regulate the response conflicts of FORs (IFOR-IFOR, EFOR-IFOR) to investigate the conflict processing among these FORs.
With respect to attention allocation, we found that our brain could efficiently allocate their attention to different FORs to influence the resposne. Specifically, the cueing effect showed that compared to the less likely target condition, in the more likely target condition, participants got a shorter response time (RT), a less error rate (ER), enhanced central P3 (396-726 ms), lower ERSP on alpha (8-13 Hz) and beta (13-20 Hz) bands, lower ITC on theta (4-8 Hz), alpha and beta bands at 400-800 time window. The time course results showed that in the cue phase, there was a longer FD and a higher FP for the FOR with high predictiveness. After target appeared, results showed a higher FP for the target FOR, no matter its predicitiveness in the cue phase. With respect to conflict processing, we found the IFOR-IFOR and EFOR-IFOR conflicts, and the interaction between them. This pattern no only showed in the RT results, but also showed on neural activities with specific conflict monitoring and shared cognitive control. ERP results showed more negative amplitudes on N2 (276-326 ms) and P3 (396-726 ms) for the incongruent conditions of these two conflicts than the congruent conditions. What’s more, there was also an interaction between them on the later P3 amplitudes (561-726 ms). Time-frequency analysis revealed that in the time window of 200-400 ms, IFOR-IFOR conflict specifically modulated power in theta bands while EFOR-IFOR conflict specifically modulated power in beta bands. However, in the time window of 400-800 ms, both conflicts modulated power in alpha and beta bands.
Moreover, computational neural network showed the cueing effect and the conflict effect on the computational cycle result similar to the RT results. Further analysis showed neurons learned through the input information to develop sparse distributed representations for task-relevant stimulus with different predictiveness. The representation with stronger predicitveness got the higher activation which might be the cause of cueing effect. Different representations shared some high sensitive neurons which might be the cause of the conflict effects. Time-course analysis showed that the activations of different representations followed by the change of the predictiveness of different representations, which fitted the human eye-tracking data well.
In sum, our results provide explanations and supports for the parallel process mechanism of multiple FORs at the representational level and the neural computational level. At the neural computational level, predictive learning, a process of spatiotemporal integration, kept the brain in a dynamic homeostatic state to fit the current input information best at each time point, drove the emergence of attention allocation and response selection at the representational level.