人们是如何记住事物的？麻省理工的科学家揭开秘密

About 50 years ago, neuroscientists discovered a type of cell in the brain's hippocampus that is associated with storing memories of specific locations. These cells are not only essential for spatial memory, but also play an important role in storing episodic memories. Although the scientific community has a deep understanding of the mechanism by which these place cells encode spatial memories, how they participate in the encoding of episodic memories remains a mystery.

大约 50 年前，神经科学家在大脑海马体中发现了一种与存储特定位置记忆相关的细胞。这些细胞不仅对空间记忆至关重要，在存储情节记忆方面同样发挥着重要作用。尽管科学界已经深入了解这些位置细胞编码空间记忆的机制，但它们如何参与情节记忆的编码仍然是一个未解之谜。

A new model developed by MIT researchers provides new insights into this question, showing that place cells can work with grid cells in the entorhinal cortex to act as a scaffold to organize memories into an interconnected sequence without spatial information.

近期，MIT 研究人员开发的一种新模型为这一问题提供了新的见解。该模型表明，位置细胞可以在没有空间信息的情况下与内嗅皮层的网格细胞协作，充当一种支架，将记忆组织成一个相互关联的序列。

"This model provides a preliminary understanding of the role of the entorhinal cortex and hippocampus in the episodic memory circuit and provides a foundation for understanding the nature of episodic memory," says Ila Fiete, a professor of brain and cognitive sciences at MIT, a member of the McGovern Institute for Brain Research, and senior author of the new study.

“这个模型初步揭示了内嗅皮层与海马体在情节记忆回路中的作用，为理解情节记忆的本质提供了基础。” MIT 大脑与认知科学系教授、麦戈文脑研究所成员、这项新研究的资深作者 Ila Fiete 表示。

This model successfully reproduces several key features of the biological memory system, including its huge storage capacity, the gradual decay of old memories, and people's ability to store large amounts of information through the "memory palace" technique.

这个模型成功再现了生物记忆系统的多个关键特征，包括巨大的存储容量、旧记忆的逐步衰退，以及人们通过“记忆宫殿”技术存储大量信息的能力。

The study was led by MIT research scientists Sarthak Chandra and Sugandha Sharma, and the results were recently published in Nature. Rishidev Chaudhuri, an assistant professor at the University of California, Davis, also participated in the study.

这项研究由MIT研究科学家Sarthak Chandra和Sugandha Sharma共同主导完成，研究成果已于近期发表在自然杂志上。加州大学戴维斯分校助理教授 Rishidev Chaudhuri 也参与了这项研究。

Memory Index

记忆的索引

To encode spatial memories, place cells in the hippocampus work closely with grid cells, a special type of neuron that fires in multiple different locations and is arranged in a regular pattern of repeating triangular geometries. Together, groups of grid cells form a triangular network that represents physical space.

为了编码空间记忆，海马体中的位置细胞与网格细胞密切合作。网格细胞是一类特殊的神经元，它们会在多个不同的位置被激活，并按照重复的三角形几何图案规律地排列在一起。网格细胞群体共同形成了一个三角形网络，用于表示物理空间。

"The same hippocampus and entorhinal circuits are used for both spatial memory and general episodic memory," Fiete explained. "So a further question we could explore is, what is the connection between spatial memory and episodic memory that allows them to share the same circuit?"

“同样的海马体和内嗅回路既用于空间记忆，也用于一般的情节记忆。” Fiete 解释道，“因此我们可以进一步探讨的问题是，空间记忆和情节记忆之间存在什么样的联系，使它们能够共享同一个回路？”

There are two main hypotheses for this overlap. The first hypothesis is that the main function of this circuit is to store spatial memory (for example, remembering where food is or where predators are, because such memories are essential for survival). In this view, the encoding of episodic memory is considered a byproduct of spatial memory.

针对这一功能重叠，科学界提出了两种主要假设。第一种假设认为，这一回路的主要功能是存储空间记忆（例如记住食物在哪里或捕食者出现的地方，因为这类记忆对于生存至关重要）。在这种观点下，情节记忆的编码被认为是空间记忆的一种副产品。

The second hypothesis, in contrast, proposes that this circuit is primarily used to store episodic memories, but can also encode spatial memories, as place is often an important component of episodic memory.

第二种假设则相反，认为这一回路主要用于存储情节记忆，同时也能编码空间记忆，因为地点通常是情节记忆中的一个重要组成部分。
原创翻译：龙腾网 https://www.ltaaa.cn 转载请注明出处

In this study, Fiete and her team proposed a third possibility: that the unique tiled structure of grid cells and their interaction with the hippocampus are equally important for both types of memory, episodic and spatial. To develop this new model, they drew on computational models developed in Fiete's lab over the past decade that simulate how grid cells efficiently encode spatial information.

在这项研究中，Fiete和她的团队提出了第三种可能性：网格细胞的独特平铺结构及其与海马体的相互作用，对情节记忆和空间记忆这两种记忆类型同样重要。为了开发这一新模型，他们借鉴了 Fiete 实验室过去十年中开发的计算模型，这些模型模拟了网格细胞如何高效编码空间信息。

"We've reached a point where I think we have a good understanding of the mechanisms of the grid cell circuit, so now it's time to explore the interactions between grid cells and larger circuits that include the hippocampus," Fiete said.

“我们已经到了这样一个阶段，我认为我们对网格细胞回路的机制有了一定程度的理解，因此现在是时候探索网格细胞与包括海马体在内的更大回路之间的相互作用了。” Fiete 说道。

In this new model, the interaction between grid cells and hippocampal cells can serve as a scaffold for storing spatial or episodic memories. Each activation pattern in the grid defines a "well," which is arranged at regular intervals. It is important to note that these wells do not directly store the content of a specific memory, but rather act as a "pointer" to a specific memory. The specific memory content is stored in the synapses between the hippocampus and the sensory cortex.

在这个新模型中，网格细胞与海马细胞的相互作用可以作为存储空间记忆或情节记忆的支架。网格中的每种激活模式定义了一个“井”，这些井按照规则的间隔排列。需要注意的是，这些井并不直接存储特定记忆的内容，而是充当指向特定记忆的“指针”。具体的记忆内容则被存储在海马体与感觉皮层之间的突触中。

When the memory is later triggered by fragmented fragments, the interaction between grid cells and hippocampal cells guides the circuit state to the nearest "well". The bottom state of these wells is connected to the corresponding part of the sensory cortex, which fills in the details of the memory. Because the sensory cortex is much larger than the hippocampus, it can store a large amount of memory information, thus compensating for the limited capacity of the hippocampus.

当记忆稍后被零散的片段触发时，网格细胞和海马细胞之间的相互作用会将回路状态引导至最近的“井”。这些井的底部状态与感觉皮层的相应部分相连，补充了记忆的细节。由于感觉皮层的规模远大于海马体，它能够存储大量的记忆信息，从而弥补海马体容量的限制。

"Conceptually, we can think of the hippocampus as a pointer network," Fiete explained. "It acts like an index, taking in partial inputs to complete a pattern and then pointing those inputs to the sensory cortex, where these memories are initially experienced and encoded. This scaffold does not contain the specific content of the memory; it only contains an index to an abstract scaffold state."

“从概念上看，我们可以将海马体视为一个指针网络。” Fiete 解释道，“它的作用类似于一个索引，通过接收部分输入完成模式补全，然后将这些输入指向感觉皮层（即这些记忆最初被体验和编码的地方），这种支架并不包含记忆的具体内容，它只包含指向抽象支架状态的索引。”

This structure enables events that occur in sequence to be lixed together: each "well" in the grid cell-hippocampal network effectively stores the information needed to activate the next "well", allowing memories to be recalled in the correct order.

这样的结构使按顺序发生的事件能够被连结在一起：网格细胞-海马网络中的每个“井”有效地存储了激活下一个“井”所需的信息，从而使得记忆能够以正确的顺序被回忆。

Modeling Memory Cliffs and Memory Palaces

建模记忆断崖与记忆宫殿

The researchers' new model reproduced several memory-related phenomena more accurately than existing models based on Hopfield networks, a type of neural network that can store and recall patterns and provided important early insights into how memories form, particularly how information is stored by strengthening connections between neurons.

研究人员的新模型在再现多个与记忆相关的现象方面比基于 Hopfield 网络的现有模型更为准确。 Hopfield 网络是一种能够存储和回忆模式的神经网络，为记忆的形成提供了早期的重要见解，尤其是关于如何通过增强神经元之间的连接来存储信息。
原创翻译：龙腾网 https://www.ltaaa.cn 转载请注明出处

However, the Hopfield network does not perfectly simulate how biological memory actually works. In this model, each memory can be recalled in great detail until the limit of storage capacity is reached. Once this limit is reached, not only is the model unable to form new memories, but even attempting to add new memories results in the complete loss of all previous memories.

然而，Hopfield 网络并不能完美模拟生物记忆的实际工作方式。在这种模型中，每个记忆都可以被细致入微地回忆，直到达到存储容量的极限。一旦达到这个极限，模型不仅无法形成新的记忆，甚至试图添加新记忆会导致所有先前记忆的完全丢失。

This "memory cliff" phenomenon does not match the actual mechanism of the biological brain. In contrast, the memory system of the biological brain is more flexible: the details of old memories will gradually blur over time, but they will not disappear completely, and new memories can be continuously stored.

这种记忆断崖的现象与生物大脑的真实机制并不相符。与之相比，生物大脑的记忆系统更加灵活：旧记忆的细节会随着时间逐渐模糊，但不会完全消失，同时可以不断存储新记忆。

The new MIT model successfully captures key insights accumulated over decades of research on grid cells and hippocampal cell recordings, primarily from studies of rodents exploring and foraging in a variety of environments. More importantly, the model also helps reveal the mechanisms behind a well-known memory strategy known as the "memory palace."

MIT 的新模型成功捕捉了数十年来通过研究网格细胞和海马细胞记录所积累的关键发现。这些记录主要来自啮齿动物在各种环境中探索和觅食行为的研究。更重要的是，该模型还帮助揭示了一种广为人知的记忆策略（即“记忆宫殿”）背后的机制。

In memory competitions, one of the tasks for contestants is to remember one or more shuffled decks of playing cards. They typically accomplish this by assigning each card to a specific location in a "memory palace." This is often a mental representation of a childhood home or other familiar place, and when the cards need to be recalled, they mentally walk through the palace, seeing the cards in turn at each location along the way. At first glance, associating cards with locations may seem to add an extra burden to the memory, but in fact, this method makes memories more solid and reliable.

在记忆力比赛中，参赛者的一项任务是记住一副或多副打乱顺序的扑克牌。他们通常通过将每张牌分配到“记忆宫殿”的特定位置来完成这一任务。记忆宫殿通常是对童年家园或其他熟悉场所的脑海再现，当需要回忆这些牌时，他们会在脑海中漫步在这个宫殿中，沿途在每个位置依次看到这些牌。乍一看，将牌与位置相关联似乎增加了额外的记忆负担，但实际上，这种方法让记忆变得更加牢固和可靠。

The MIT team's computational model was able to perform tasks similar to the "Memory Palace" very well, suggesting that this memory strategy actually exploits the natural mechanism of the memory circuit. By associating input with scaffolds in the hippocampus and combining long-term memories stored in the sensory cortex as new scaffolds, this approach significantly expands the capacity of the memory system, allowing humans to store and recall more sequence information than usual.

MIT 团队的计算模型能够出色地完成类似“记忆宫殿”的任务，这表明这种记忆策略实际上是利用了记忆回路的自然机制。通过将输入与海马体中的支架关联起来，并结合长期存储在感觉皮层中的记忆作为新的支架，这种方法显著扩展了记忆系统的容量，使得人类能够存储和回忆比平时更多的序列信息。

Next, the researchers plan to further explore several important issues based on this model, such as how episodic memory is converted into semantic memory in the cortex, that is, factual memory that is independent of the specific context when it is acquired, such as "Paris is the capital of France"; how events are defined and organized; and how to integrate brain-like memory models into modern machine learning systems to enhance the memory and learning capabilities of artificial intelligence.

接下来，研究人员计划基于这一模型进一步探索几个重要问题，比如情节记忆如何转化为皮层中的语义记忆，即与获取时的具体情境无关的事实性记忆，例如巴黎是法国的首都; 事件如何被定义和组织; 以及如何将类脑记忆模型整合到现代机器学习系统中，从而提升人工智能的记忆和学习能力。