Neural Mechanisms of Learning Relevance in Multidimensional Environments

多维环境中学习相关性的神经机制

• 批准号: 10380142
• 负责人: Thilo Womelsdorf
• 金额: $ 62.91万
• 依托单位: VANDERBILT UNIVERSITY
• 依托单位国家: 美国
• 财政年份: 2021
• 资助国家: 美国
• 起止时间: 2021-04-01 至 2026-01-31
• 项目状态: 未结题
• 来源: https://reporter.nih.gov/project-details/10380142
• 关键词: Address; Animals; Anterior; Area; Attention; Behavioral; Behavioral Model; Brain; Brain region; Cells; Code; Cognitive; Complex; Corpus striatum structure; Data; Decision Making; Dimensions; Dorsal; Electrophysiology (science); Environment; Focused Ultrasound; Future; Hybrids; Interneurons; Intervention; Learning; Lesion; Mediating; Modeling; Monkeys; Neurons; Outcome; Performance; Population; Prefrontal Cortex; Primates; Probability; Process; Protocols documentation; Psychological reinforcement; Regulation; Research; Rewards; Role; Short-Term Memory; Speed; System; Testing; Uncertainty; Update; Visual; attentional bias; base; cell type; cingulate cortex; cognitive change; expectation; flexibility; high dimensionality; high reward; learning strategy; neural circuit; neuromechanism; nonhuman primate; recruit; relating to nervous system; tool

PROJECT SUMMARY / ABSTRACT This proposal investigates in the nonhuman primate how attentional load changes the behavioral and neural strategies for flexibly learning object relevance. High attentional load characterizes real-world learning scenarios with multiple, multidimensional objects. Evidence suggests that the neural mechanisms underlying learning during high attentional load fundamentally differs from neural mechanisms used to learn under low load. Our proposal elucidates how learning at increasing attentional load (1) changes the cognitive subcomponent processes used to succeed learning, (2) changes which brain areas are used to flexibly learn, and (3) recruits additional neural circuit mechanisms to realize fast adjustments. First, we will address the specific behavioral subcomponent processes used for learning the relevance of objects in environments with increasing number of visual feature dimensions reflecting increasing attentional load. Simple learning can be achieved efficiently with a hybrid mechanism that uses working memory (WM) of recently rewarded objects to guide future choices together with slower reinforcement learning (RL) for updating longer- term value expectations. When attentional load increases working memory breaks down, and efficient learners flexibly adjust their exploration rates and attentional prioritization to speed up reinforcement learning. Our proposal quantifies these changing learning strategies with multi-component WM-RL modeling. Second, while subjects learn with varying strategies which features to use for making a decision, we will test the causal role of three brain regions implicated to realize the respective learning mechanisms. We use transcranial focused ultrasound stimulation to induce transient, fully reversible lesions allowing to functionally disrupt confined neuronal ensembles. With this tool we elucidate the hypothesized contributions of ventrolateral prefrontal cortex to learning using fast working memory of rewarded objects, the contribution of the anterior cingulate cortex in adjusting exploration strategies and the contribution of the anterior striatum for attentional biasing of slower reinforcement learning of the highest reward-value object within a complex, multidimensional feature space. Third, our project elucidates how the local circuits in each of the three brain areas contribute to successful learning with varying strategies. We use massively parallel recordings of single neuron activity in ventrolateral prefrontal cortex, anterior cingulate cortex, and anterior striatum to extract those cell classes whose firing encodes the key learning variables. We expect that subclasses of interneurons maximally correlate their firing only during those periods when the area specific learning strategy is realized. This approach pinpoints the cell classes that maximally correlate with choice probabilities, prediction errors, working memory, and exploration rates when subjects adjust their learning strategies to successfully learn the relevance of objects with real-world complexity.

项目概要/摘要 该提案研究了非人类灵长类动物的注意力负荷如何改变行为和神经 灵活学习对象相关性的策略。高注意力负荷是现实世界学习场景的特征 与多个、多维的对象。有证据表明，学习背后的神经机制 在高注意力负荷下学习的神经机制与在低负荷下学习的神经机制根本不同。我们的 提案阐明了增加注意力负荷时的学习 (1) 如何改变认知子成分 用于成功学习的过程，（2）改变用于灵活学习的大脑区域，以及（3）招募 额外的神经回路机制以实现快速调节。 首先，我们将讨论用于学习对象相关性的特定行为子组件过程 在视觉特征维度数量不断增加的环境中，反映了注意力负荷的增加。 通过使用最近的工作记忆（WM）的混合机制可以有效地实现简单的学习 奖励对象与较慢的强化学习（RL）一起指导未来的选择，以更新更长的时间 术语价值期望。当注意力负荷增加时，工作记忆就会崩溃，高效的学习者 灵活调整他们的探索率和注意力优先顺序，以加速强化学习。我们的 提案通过多组件 WM-RL 建模来量化这些不断变化的学习策略。 其次，当受试者使用不同的策略学习哪些特征用于做出决策时，我们将测试 三个大脑区域的因果作用涉及实现各自的学习机制。我们使用经颅 聚焦超声刺激可诱导短暂的、完全可逆的损伤，从而破坏受限的功能 神经元集合。通过这个工具，我们阐明了腹外侧前额皮质的假设贡献 对于使用奖励对象的快速工作记忆进行学习，前扣带皮层的贡献 调整探索策略和前纹状体对慢速注意力偏差的贡献 对复杂的多维特征空间中最高奖励值对象的强化学习。 第三，我们的项目阐明了三个大脑区域中每个区域的局部回路如何有助于成功 使用不同的策略进行学习。我们使用腹外侧单个神经元活动的大规模并行记录 前额皮质、前扣带皮层和前纹状体以提取那些放电的细胞类别 对关键学习变量进行编码。我们期望中间神经元的子类与其放电最大程度地相关 仅在实现特定领域学习策略的时期。这种方法可以精确定位细胞 与选择概率、预测误差、工作记忆和探索最大程度相关的类别 受试者调整学习策略以成功学习物体与现实世界的相关性的比率 复杂。