Computational modeling of generative episodic memory

尽管大量的实验和概念研究表明情节记忆是生成性的，但计算模型几乎完全采用存储的观点。在这个项目中，我们开发了一个个人经历情节的编码和检索的生成模型，该模型描述了海马体和新皮层之间的相互作用。该模型包括：(A)一个层次结构的感知-语义网络，该网络将感知的图像逐渐转换为更具语义的表征；(B)一个语义网络，它能够在递归过程中以可信的方式补充不完整的语义表征。前者由“矢量量化变分自动编码器(VQ-VAE)”实现，后者由“像素卷积神经网络(PixelCNN)”实现。当一集被编码时，VQ-VAE首先将其转换为语义表示，然后存储其中的一部分。当注意力较高时，大部分会被储存；当注意力较低时，只会储存一小部分。在召回过程中，这一部分被再次读出，并由PixelCNN可信地完成。到目前为止，我们使用不同背景上的手写数字的单个图像作为情节；数字表示不同变体中的对象，背景表示上下文，例如可以找到对象的房间。对象或数字最好是在特定的上下文中或在某些背景的前面找到的，例如厨房中的烤面包机(一致的上下文)，而不是在浴室(不一致的上下文)中，或者在我们的模拟中，背景前面的‘2’是三角形而不是正方形。我们已经用该模型再现了以下实验结果：(I)更高的注意力可以改善情景记忆；(Ii)在一致的语境中的物体比在不一致的语境中记住得更好；(Iii)如果一个物体没有记住正确的语境，通常至少会记住一个语义上一致的语境。例如，我们不喜欢回忆那些让我们感到尴尬的情况，我们喜欢让我们的记忆更符合我们在回忆中对自己的印象。相反，我们的记忆自然会影响我们的自我形象。我们还知道，我们的情景记忆可以通过社会互动来改变。特别是，当我们感觉到与我们的互动伙伴联系在一起时，我们倾向于将记忆与他们的观点保持一致。这些方面是我们与思考自我模型的哲学家和正在进行社交互动对记忆影响的实验的心理学家合作，对我们的模型进行进一步研究的主题。

Despite the large number of experimental and conceptual studies that have suggested that episodic memory is generative, computational models almost exclusively adopt the storage view. In this project, we develop a generative model for the encoding and retrieval of personally experienced episodes, which describes the interplay between hippocampus and neocortex.The model consists of (a) a perceptual-semantic network that is hierarchically structured and gradually transforms perceived images into a more semantic representation, and (b) a semantic network that is able to complement incomplete semantic representations in a plausible way in a recurrent process. The former is realized by a 'vector quantized variational autoencoder (VQ-VAE)', the latter by a 'pixel convolutional neural network (PixelCNN)'. When an episode is encoded, the VQ-VAE first converts it into a semantic representation, part of which is then stored. When attention is high, a large part is stored; when attention is low, only a small part is stored. During recall, this part is read out again and plausibly completed by the PixelCNN. The VQ-VAE can then be applied backwards and reconstruct a concrete episode from the complete semantic representation.So far, we use single images of handwritten digits on different backgrounds as episodes; the digits represent objects in different variants, the backgrounds represent the context, e.g. the room in which the object can be found. Objects or digits are preferably found in certain contexts or in front of certain backgrounds, e.g. a toaster in the kitchen (congruent context) and not in the bathroom (incongruent context) or in our simulation a '2' in front of a background with triangles and not squares. We have already reproduced the following experimental results with the model: (i) higher attention improves episodic memory, (ii) objects in congruent context are better remembered than in incongruent context, and (iii) if the correct context is not remembered for an object, at least a semantically congruent context is usually remembered.Episodic memory is not reliable and can be modified by many influences. For example, we do not like to remember situations that were embarrassing to us, and we like to bring our memories more in line with the image we have of ourselves in retrospect. Conversely, our memories naturally influence our self-image. It is also known that our episodic memory can be altered by social interaction. In particular, we tend to align memories with opinions of our interaction partners when we feel connected to them. These aspects are the subject of further research on our model in cooperation with philosophers who are thinking about the self-model and with psychologists who are doing experiments on the influence of social interaction on memory.