Information retrieval by a neural-network system with continuous attractors

长期以来，人们一直假设神经网络系统中的信息检索是由具有离散不动点吸引子的动力系统来描述的。然而，分级持续活动的神经生理学发现及其神经机制的计算模型的证据表明，从大脑的长期记忆中提取短期记忆更有可能被描述为具有持续依赖于初始状态的固定点吸引子的动力学(例如，连续吸引子动力学)。在心理学中，人们普遍认为长时记忆是在一个网络结构(如语义网络)中存档的。在这里，我们提出了从各种现实世界的复杂网络(WWW/Internet、科学论文之间的引文、人际网络、社会网络、基因/生化反应网络等)中进行信息检索。通过连续吸引子动力学，类比从长期记忆的大型网络中提取短期记忆。对于给定的COM…更复杂的网络，每个节点对应一个具有滞后输入输出关系的神经元，每个链接对应一个突触连接。用户想要知道的(即，用户的“询问”)被编码在神经元激活模式的初始状态中。假定每个神经元的滞后特性对于产生稳健的连续吸引子是必不可少的。作为连续吸引子获得的激活模式表示对查询的“答案”。通过将这种信息检索算法应用于科学文章的引文网络(300,000个神经科学提供的科学引文索引扩展，Thomson Science，经允许)，我们确认，响应于给定的查询，充分地提取了一组相关文档。例如，对于“分级持续活动和神经整合者”的查询，通过可视化提取的文档及其之间的引用关系，可以理解哪些文章是主从性的，哪些文章之间的关系是支流。为了阐明真实神经元是否具有像所提出的算法中假设的那样的滞后特征，我们分析了在GO/NO-GO辨别任务中从猴子扣带回皮质记录的分级活动。我们发现，放电频率分布呈现明显的双峰性，这与具有滞后特性的神经元的理论预测是一致的。研究还表明，具有滞后特性的神经元递归网络可以很好地复制实验中观察到的分级活动特征。这些结果表明，所提出的算法的信息检索非常类似于从大脑的长期记忆中提取短期记忆。据我们所知，这是第一个从真实大脑中推断出非平凡的、实用的信息处理算法的成功例子。较少

It has long been hypothesized that information retrieval in neural-network systems is described by dynamical systems with discrete fixed-point attractors. However, evidence from neurophysiological findings of graded persistent activity and computational modeling of its neural mechanisms suggests that retrieval of short-term memory from long-term memory in the brain is more likely to be described by dynamics with fixed-point attractors that continuously depend on the initial state (say, continuous-attractor dynamics). In psychology, it has been generally considered that long-term memory is archived in a network structure (e.g. semantic network). Here we propose information retrieval from a variety of real-world complex networks (WWW/internet, citation between scientific articles, human network, social network, gene/biochemical-reaction network, etc.) by continuous-attractor dynamics, by analogy with retrieval of short-term memory from a large network of long-term memory. For a given com … More plex network, a neuron with hysteretic input/output relation corresponds with each node, and a synaptic connection with each link. What a user wants to know (i.e. user's "query") is encoded in an initial state of the activation pattern of the neurons. The hysteretic characteristics assumed for each neuron are essential for producing robust continuous attractors. An activation pattern obtained as a continuous attractor represents an "answer" to the query. By applying this information-retrieval algorithm to a citation network of scientific articles (300,000 neuroscience provided Science Citation Index Expanded, Thomson Scientific, with permission), we confirmed that, in response to a given query, a set of relevant documents were adequately extracted. For instance, for a query "graded persistent activity and neural integrator", by visualizing the extracted documents and citation relations between them, one can comprehend which articles are principal or accessory and which relations between articles are mainstreams of tributaries. To elucidate whether real neurons have hysteretic characteristics such as those hypothesized in the proposed algorithm, we analyzed graded activity recorded from the monkey cingulate cortex during Go/No-go discrimination task. We found that the firing-rate distribution shows clear bimodality, which is consistent with the theoretical prediction for neurons with hysteretic characteristics. It was also demonstrated that a recurrent network of neurons with hysteretic characteristics could well replicate experimentally observed features of graded activity. These results suggest that information retrieval by the proposed algorithm is quite analogous to short-term memory retrieval from long-term memory in the brain. To our knowledge, this is the first successful example to infer non-trivial, practically useful information-processing algorithm from real brain. Less