恐惧学习是关乎生物体生存最重要的学习形式，对于拥有优秀学习能力的人类而言，少量的恐惧学习足以通过泛化帮助我们规避相似的潜在威胁，提高生存率。但恐惧泛化是一把双刃剑，帮助我们生存的同时，也可能极大影响我们的生活质量与心理健康。过度泛化使个体极力躲避任何潜在威胁，耗费大量不必要时间精力的同时常常感到情绪不良，变得社会适应困难。所以，关于恐惧学习与泛化的研究于人类而言十分重要，未来有可能启示我们如何把握好恐惧学习与泛化的平衡。

相比传统一维恐惧学习与泛化研究，本研究选取由电击概率和主观疼痛强度构成的风险作为恐惧刺激进行更具生态效度的二维恐惧学习与泛化研究。由于近年来，网格细胞被广泛发现用以编码各类二维空间，且支持概化信息、泛化推理，灵活决策等功能，我们以网格细胞作为切入点深入研究二维恐惧学习与泛化的底层神经机制。我们认为，在基于风险的二维恐惧学习后，个体形成由网格细胞参与编码的二维恐惧空间，恐惧泛化模式与该空间内部结构有关。

为了验证该假设，我们分别设计了行为与磁共振实验，具体过程与结果如下：

在行为实验中，我们想探究二维恐惧学习后，个体是否会形成特异性二维恐惧空间，我们用基于风险的恐惧反应模型来抽象表达该空间形态。实验过程主要让被试学习9种中性线索与真实电击风险的关系，随后进行泛化探测。我们假设，中性线索在恐惧学习后，个体即可形成中性线索与电击风险关系的恐惧认识，这种认识可用基于风险的恐惧反应模型来表征，已习得恐惧的中性线索会根据模型将恐惧泛化至未习得恐惧的中性线索。结果发现，基于风险的恐惧反应模型对恐惧学习前的中性线索恐惧量预测的偏差显著大于恐惧学习后的，而在对恐惧学习后学习条件与泛化条件间中性线索恐惧量进行预测时，则发现两者的偏差无显著差异。换言之，个体在进行风险恐惧学习后形成了特异性二维恐惧空间，且恐惧泛化的模式根据该空间模式进行。

在磁共振实验中，我们想探究个体恐惧学习后形成的特异性二维恐惧空间是否由网格细胞编码，其编码方式如何指导恐惧泛化进行。实验过程主要是让被试在磁共振中完成恐惧泛化导航任务，该任务包括“恐惧导航”、“恐惧定位”、“恐惧评分”三个阶段，其中最重要的阶段是“恐惧导航”阶段，该阶段要求被试观看中性线索在2s内的变化过程。对该阶段的磁共振数据进行网格细胞信号分析，结果发现，右侧内嗅皮层（EC）存在六周期信号激活，且在右侧脑岛（INS），右侧眶额皮层（OFC）和左侧杏仁核发现与右侧内嗅皮层网格方向一致的六周期信号。六周期信号强度与行为指标关联分析发现，右侧OFC的六周期信号强度与恐惧泛化导航任务时“恐惧比较”任务的准确率呈高度正相关，左侧杏仁核的六周期信号强度则与准确率呈中度负相关。

我们推测，右侧内嗅皮层系统杏仁核（恐惧学习与反应的关键脑区）、脑岛（恐惧泛化相关脑区）、眶额皮层（价值评估相关脑区）共同编码了一个特异性二维恐惧空间以帮助个体进行基于风险的恐惧泛化，协同模式影响恐惧空间编码效果。

我们的行为与核磁实验结果共同支持了我们的研究假设，个体在恐惧学习过程中，逐渐形成了由网格细胞编码的特异性二维恐惧空间，并因此使习得的恐惧以某种特定的模式进行泛化。

Fear learning is the most important form of
learning related to the survival of organisms. For human with excellent
learning ability, a small amount of fear learning is enough to help us avoid
similar potential threats and improve survival through generalization. But fear
generalization is a double-edged sword. While helping us survive, it may also
greatly affect our quality of life and mental health. Overgeneralization makes
individuals try to avoid any potential threats, spend a lot of unnecessary time
and energy, and often feel bad mood and become difficult to adapt to the
society. Therefore, research on fear learning and generalization is very
important for human, and it may enlighten us how to grasp the balance between
fear learning and generalization in the future.

Compared with the traditional one-dimensional
fear learning and generalization research, this study selected the risk
composed of electric shock probability and subjective 疼痛 intensity as the fear stimulus to
conduct a more ecologically valid two-dimensional fear learning and
generalization research. In recent years grid cells have been widely discovered
to encode various two-dimensional spaces and support functions such as generalized
information, generalized reasoning, and flexible decision-making. Consequently
we use grid cells as an entry point to deeply study the underlying neural
mechanisms of two-dimensional fear learning and generalization. We argue that
after risk-based two-dimensional fear learning, individuals form a
two-dimensional fear space encoded by grid cells, and that fear generalization
patterns are related to the internal structure of this space.

In order to verify this hypothesis, we designed
behavioral experiment and fMRI experiment respectively. The specific process
and results are as follows:

In behavioral experiment, we want to explore
whether individuals will form a specific two-dimensional fear space after
two-dimensional fear learning, and we use a risk-based fear response model to
abstractly express this spatial form. The experimental process mainly allows
the participants to learn the relationship between 9 neutral cues and the real
risk of electric shock, and then conduct generalization detection. We
hypothesize that after fear learning of neutral cues, individuals can form a
fear cognition about the relationship between neutral cues and electric shock
risk. This cognition can be represented by a risk-based fear response model.
Neutral cues that generalize fear to unlearned fear. The results showed that
the risk-based fear response model had a significantly greater bias in
predicting the amount of neutral cue fear before fear learning than after fear
learning. However, when predicting the amount of neutral cued fear between the
learning condition and the generalization condition after fear learning, no
significant difference was found between the two deviations. In other words,
individuals form a specific two-dimensional fear space after performing risk
fear learning, and the pattern of fear generalization is carried out according
to this spatial pattern.

In the fMRI experiment, we wanted to explore
whether the specific two-dimensional fear space formed after individual fear
learning is encoded by grid cells, and how the encoding method guides fear
generalization. The experimental process is mainly to let the participants
complete the fear generalization navigation task in the scanner. The task includes three stages:
"fear navigation", "fear localization" and "fear
scoring", of which the most important stage is the "fear
navigation" stage , this stage requires subjects to watch the change
process of neutral cues within 2s. Grid cell signal analysis was performed on
the fMRI data at this stage, and it was found that there was six-fold signal
activation in the right
entorhinal cortex, and a six-period signal with  grid
orientation consistent with the
right entorhinal cortex was found in the
right insula, the right orbitofrontal cortex and the left amygdala. The correlation analysis between the six-fold signal strength and
behavioral indicators found that the six-fold signal strength of the right OFC
was highly positively correlated with the accuracy of the "fear
comparison" task during the fear generalization navigation task, while the
six-fold signal strength of the left amygdala was significantly correlated with
the accuracy of the "fear comparison" task during the fear
generalization navigation task. Accuracy was moderately negatively correlated.

We speculate that the right entorhinal cortex
combined with the amygdala (a key brain area for fear learning and response),
the insula (a brain area related to fear generalization), and the orbitofrontal
cortex (a brain area related to value assessment) jointly encode a specific
Dimension fear space to help individuals with risk-based fear generalization.

Our behavioral and fMRI experimental results
together support our research hypothesis. During the process of fear learning,
individuals gradually form a specific two-dimensional fear space encoded by
grid cells, and thus make the learned fear generalize in a specific mode.