RI: Small: Generation and modulation of variability in trial and error learning

RI：小：试错学习中变异性的生成和调节

• 批准号: 1421245
• 负责人: Adrienne Fairhall
• 金额: $ 45.95万
• 依托单位: University of Washington
• 依托单位国家: 美国
• 项目类别: Standard Grant
• 财政年份: 2014
• 资助国家: 美国
• 起止时间: 2014-09-01 至 2018-08-31
• 项目状态: 已结题
• 来源: https://www.nsf.gov/awardsearch/showAward?AWD_ID=1421245&HistoricalAwards=false
• 关键词: RI; Small; Generation; modulation; variability

Reinforcement learning is a powerful learning strategy in which random perturbations are made to an output and changes that lead to more successful outputs are reinforced. Animals learn to perform highly skilled tasks by trial and error, but how biological neural networks generate variability from trial to trial is not known. An ideal model for examining this relationship between variability and learning can be found in song learning by birds. Birds learn their songs by trial and error, and are able to adjust their songs into adulthood. In addition, the basic neural architecture of the learning circuit has been mapped out; some effects of behavioral context on variability are also known. How variability is generated, and how it is controlled, can thus be studied mechanistically.This project will study two candidate mechanisms for the generation and control of variability: chaotic dynamics in a cortex-like area, and dynamical modulation of synchrony in a basal ganglia circuit. While generic neural networks are chaotic, this project will explore how such dynamics are modified in a cortical network of biophysically plausible neurons with structured connectivity based on the known topography of the learning circuit. Furthermore, the birdsong cortical area receives inputs generated by basal ganglia. How might this input influence the variability of the output? It is proposed that the ability of activity in basal ganglia to affect cortex may depend on the synchrony of basal ganglia outputs. The studies will be constrained by experimental data from these candidate areas through collaborations with two laboratories. The work will contribute to a general understanding of variability in biological learning, and may suggest strategies for structuring noisy input during motor learning that can lead to efficient outcomes for robotics applications. Further, Dr Fairhall is developing materials for broad outreach programs that include a Massive Online Open Course to introduce concepts in computational neuroscience to a very diverse audience, and a workshop for middle school girls interested in mathematics applied to the life sciences.

强化学习是一种强大的学习策略，其中对输出进行随机扰动，并加强导致更成功输出的变化。动物通过反复试验来学习执行高技能任务，但生物神经网络如何在一次又一次试验中产生可变性尚不清楚。在鸟类的鸣叫学习中可以找到一个理想的模型来研究这种变化和学习之间的关系。鸟类通过反复试验来学习它们的歌声，并且能够在成年后调整它们的歌声。此外，还绘制了学习回路的基本神经结构；行为环境对变异性的一些影响也是已知的。因此，可变性是如何产生的，以及它是如何被控制的，可以从机械上进行研究。该项目将研究两种可变性产生和控制的候选机制：类皮层区域的混沌动力学和基底节区电路同步的动态调制。虽然一般的神经网络是混沌的，但该项目将探索如何在生物物理上合理的神经元皮质网络中修改这种动态，这些神经元具有基于已知学习电路地形的结构化连接。此外，鸟鸣皮质区接受基底神经节产生的输入。这种输入如何影响输出的可变性？有人提出，基底神经节活动影响皮层的能力可能取决于基底神经节输出的同步性。通过与两个实验室合作，这些研究将受到来自这些候选领域的实验数据的限制。这项工作将有助于对生物学习变异性的一般理解，并可能提出在运动学习过程中构建噪声输入的策略，从而为机器人应用带来有效的结果。此外，费尔霍尔博士正在为广泛的推广项目开发材料，其中包括一个大型在线开放课程，向非常多样化的受众介绍计算神经科学的概念，以及一个为对数学应用于生命科学感兴趣的中学女生举办的研讨会。