Neuromodulatory-prefrontal interactions in primates

我们的目标是了解两个大脑系统如何合作和竞争来指导行为。其中一个系统由上行神经调节系统(ANS)组成。这些细胞由脑细胞(神经元)组成的核组成，这些细胞在大脑中发送投射，对行为产生广泛的影响。另一个大脑系统是前额叶和前扣带回(PFC/ACC)。虽然ANS存在于所有哺乳动物和许多其他动物中，但PFC/ACC是唯一专门针对灵长类动物的。我们假设ANS代表着有机体环境的关键特征，这些特征指导着行为的基本方面。例如，一个核团(中缝核团)可能会编码环境的平均良好程度：它产生回报的速度是高还是低？如果有机体处于一个回报很高的环境中，那么一切可能都很好，但如果不是这样，那么可能是时候寻找更好的替代方案了。其他ANS核可能编码了有机体对这种估计的不确定性。如果有机体的估计非常不确定，那么环境可能正在变化，有机体需要寻找更多的信息来建立对情况的更好估计。与ANS一样，PFC/ACC也代表了关于环境的信息，如奖励丰富度和不确定性，以指导战略行为。那么，是什么使PFC/ACC能够支持我们在灵长类动物身上观察到的复杂行为？它如何与ANS交互？我们将测试几个想法。一个核心观点是，与ANS中的信息相比，PFC/ACC中的信息要丰富得多，或者说，信息的维度更高。例如，PFC/ACC可能包含有关环境中所有可用选项的价值的更具体信息。它可以对信息各组成部分之间的关系进行编码。这可以以更复杂的方式指导行为，当我们比较PFC/ACC和ANS活动时，这一点会很明显。需要几个大脑活动测量。它们是在一种名为猕猴的灵长类动物中制造的。一种方法是用磁共振成像(MRI)扫描仪测量活动。至关重要的是，这同时告诉我们PFC/ACC和ANS的活动，以便我们可以比较它们并研究它们的相互作用。它还使人们能够与人类核磁共振研究建立联系。我们将利用我们最近开发的磁共振记录PFC/ACC和ANS的协议，而动物参与了丰富的行为曲目。第二种方法涉及电极记录大脑实际计算单元--神经元--在毫秒级时间尺度上的活动。我们将使用最新的电极同时记录数十个甚至数百个神经元。这使我们能够研究PFC/ACC中丰富的高维信息编码。最后，我们将使用一种技术，即经颅超声刺激(TUS)，它以一种相对非侵入性的方式暂时中断活动。重要的是，这种方法确定了一个大脑区域的活动如何导致其他区域的活动，并最终导致行为。我们是世界上极少数能够进行这些研究的研究团队之一。我们相信，发展这样的理解对人工智能(AI)代理来说将变得重要。有一些精确的数学方法来描述奖励信息，并在行为指导中对其进行最佳使用。类似于在人工智能中观察到的描述已经支撑了最先进的人工智能算法，这些算法通过反复试验来学习行为。这些不需要程序员显式地编程代理应该如何行为，而是指定代理的高级目标：然后代理了解哪些操作可以实现这一目标。有可能开发下一代人工智能算法，通过像灵长类动物那样更多地学习来实现他们的目标；PFC/ACC可能允许灵长类动物从多个例子中提取信息，学习环境的结构，并根据个人观察执行快速推理，这是人工智能代理目前无法做到的。

We aim to understand how two brain systems cooperate and compete to guide behaviour. One system consists of the ascending neuromodulatory systems (ANS). These comprise groups, called nuclei, of brain cells (neurons) that send projections across the brain to exert broad influences over behaviour. The other brain system is prefrontal and anterior cingulate cortex (PFC/ACC). Whereas ANS are present in all mammals and many other animals, PFC/ACC is uniquely specialized in primates.We hypothesize that ANS represent key features of an organism's environment that guide fundamental aspects of behaviour. For example, one nucleus (the raphe nucleus) may encode how good the environment is on average: does it yield rewards at a high or low rate? If the organism is in a highly rewarding environment then all may be well, but if not, then it may be time to seek a better alternative. Other ANS nuclei may encode the organism's uncertainty about such estimates. If the organism's estimates are very uncertain then the environment may be changing, and the organism needs to seek more information to establish a better estimate of the situation.Like ANS, PFC/ACC also represents information about the environment such as reward richness and uncertainty to guide strategic behaviours. What then enables PFC/ACC to support the sophisticated behaviours we observe in primates? How does it interact with ANS? We will test several ideas. One central idea is that information in PFC/ACC is much richer, or 'high dimensional', than information in ANS. For example, PFC/ACC might hold much more specific information about the value of all choices available in the environment. It may encode relationships between component pieces of information. This could guide behaviour in more sophisticated ways and this would be apparent when we compare PFC/ACC and ANS activity.Several brain activity measurements are required. They are made in a primate called a macaque. One approach measures activity with a magnetic resonance imaging (MRI) scanner. Crucially, this simultaneously tells us about activity in PFC/ACC and ANS so that we can compare them and study their interactions. It also enables links to be drawn with human MRI studies. We will exploit our recently developed protocols for MRI recording of PFC/ACC and ANS while animals engage in a rich repertoire of behaviour. A second approach involves electrodes recording activity from the actual computing units of the brain - the neurons - on the millisecond time scale on which they operate. We will use the latest electrodes to record many tens or even hundreds of neurons simultaneously. This allows us to study rich, 'high dimensional' information encoding in PFC/ACC. Finally, we will use a technique, transcranial ultrasound stimulation (TUS), which transiently disrupts activity in a comparatively non-invasive way. Importantly, this approach establishes how activity in one brain region leads to activity elsewhere and ultimately causes behaviour. We are one of very few research teams in the world that can undertake these studies.We believe that developing such an understanding will become important for artificially intelligent (AI) agents. There are precise, mathematical ways to describe reward information and its optimal use in behavioural guidance. Descriptions resembling observations made in ANS already underpin state-of-the-art AI algorithms that learn behaviour via trial-and-error. These do not require the programmer to explicitly program how the agent should behave, but instead specify the agent's high-level goal: the agent then learns what actions achieve this goal. It may be possible to develop the next generation of AI algorithms to achieve their goals by learning more like primates do; PFC/ACC may allow primates to abstract information away from multiple examples, learn the structure of environments, and perform rapid inferences based on individual observations, in a way that AI agents currently cannot.