收录解读
这篇 Nature 论文研究 dmPFC 如何把 learned cues 的 value、salience 和 valence 组织成可读出的神经几何，而不是只报告单个变量的编码相关性。
作者在自由活动小鼠中用 calcium imaging 记录 dmPFC 单神经元群体，让动物区分预测 reward 或 punishment 的刺激，从而把 salience、valence 和 value 信息拆开分析。
核心结果是 dmPFC 群体主要编码 learned stimuli 的 appetitive/aversive value，并且部分神经元把 valence 与 salience 放在近似正交的信息轴上；这些几何结构会随 associative learning 形成并预测趋近或回避行为。
它值得正式收录，因为它给 AI-relevant neuroscience 一个清晰机制：价值、显著性和效价可以通过低维、近正交的 population geometry 实现可组合行为控制，对 agent reward representation、motivated behavior、state abstraction 和安全相关趋避信号建模都有概念外溢。
原始摘要与中文对照
中文对照翻译
学习线索的前额神经几何结构引导动机行为。动物不断评估周围环境，以决定是接近奖励机会还是避免潜在威胁。为环境刺激赋予适当的重要性不仅对生存至关重要，而且是包括人类在内的跨物种共享的复杂形式的目标导向行为的基础。因此，理解大脑如何将这些感觉线索转化为动机行为是神经科学和心理学的核心。背内侧前额叶皮层 (dmPFC) 是连接相关环境刺激与目标导向行为的关键结构。显著性、效价和价值是定义刺激相关性的关键维度，但dmPFC如何处理和组织这些维度以驱动动机行为仍不清楚。在本研究中，我们使用钙成像技术监测了自由活动的雄性小鼠dmPFC中的单神经元群体，同时小鼠区分预测不同奖励或惩罚结果的刺激，这使得我们能够前所未有地分离显著性、效价和价值信息。我们发现dmPFC神经元群体主要编码学习刺激的趋近性和规避性价值，并且亚群体沿着正交信息轴编码效价和显著性。我们的结果强调了在dmPFC网络内的联想学习过程中，刺激的价值、显著性和效价的并发多方面群体编码，使得dmPFC神经元表征的几何结构动态地塑造了趋近性和规避性动机行为。
原始摘要
Animals continuously evaluate their surroundings to decide whether to approach rewarding opportunities or avoid potential threats. Assigning the appropriate importance to environmental stimuli is not only crucial for survival but also underlies complex forms of goal-directed behaviour that are shared across species, including humans . Understanding how the brain translates such sensory cues into motivated behaviours is, therefore, central to neuroscience and psychology. The dorsomedial prefrontal cortex (dmPFC) is a critical structure that bridges relevant environmental stimuli to goal-directed behaviour. Salience, valence and value are key dimensions defining stimulus relevance, but how the dmPFC processes and organizes such dimensions to drive motivated behaviour remains unclear. Here we monitored single-neuron populations in the dmPFC using calcium imaging in freely moving male mice while discriminating between stimuli predicting different reward or punishment outcomes, which enabled an unprecedented dissociation of salience, valence and value information. We found that dmPFC populations primarily encode appetitive and aversive values of learned stimuli and that subpopulations encode valence and salience along orthogonal information axes. Our results highlight a concurrent multifaceted population coding of value, salience and valence of stimuli during associative learning within dmPFC networks, such that the geometry of dmPFC neuronal representations dynamically shapes appetitive and aversive motivated behaviours.