果蝇运动目标检测环路机制的解析与计算仿真研究

Researchers in brain science and brain-like intelligence research are being encouraged to investigate the human intelligence so as to develop wiser and useful machines (MIT IQ, 2018). We are facing a new storm of the ‘struggle for survival’ in the field of artificial intelligence. Instructively, a recent study solved the nearest-neighbors search problem by emulating the approach of the olfactory circuit to identify individual odors in Drosophila (Science, 2017). Today, one of the most important advances in neurobiology recently reveal an elaborate internal structure of the neural circuit corresponding to elementary motion detector (EMD), showing amazing similarities between the mouse retina and the fly optic lobe (Nature Neuroscience, 2015). The mechanism underlying feature extraction and integration in the visual circuits postsynaptic to the EMD, however, is poorly understood. In this project we focus on how moving targets amongst visual clutter are discerned by flying fruit flies. By combining a computational modelling approach with quantitative behavioural analysis and the rich genetic tools available to Drosophila, our project mainly addresses the following three questions. First, what is the neural underpinning of the small-field system which is most sensitive to the motion of small targets? Second, whether the small-field system is independent of the large-field system which is most sensitive to coherent background motion? Finally, besides the changes in local luminance, does the position system also detect other higher-order visual features? The results are expected to reveal computation principles at neuronal and circuital levels underlying motion processing. The results are also expected to shed light on engineering technology on moving target detection.

脑科学和类脑智能研究的重要趋势是“了解人类智力进而开发智能机器”（美国MIT IQ计划，2018）。我们正面临人工智能革命新一轮的“生存竟争”风暴。具有启发意义的是，利用果蝇嗅觉编码原理解决了最近邻搜索算法问题（Science，2017）。今天神经科学重要进展之一，是绘制了从果蝇视网膜到髓质这一与初级运动检测器对应环路的神经网络结构，证明果蝇和小鼠的运动视觉计算惊人相似（Nature Neurocience, 2015）。然而，对初级运动检测器突触后的高级视叶区仍知之甚少。本课题聚焦果蝇复杂背景下的目标检测，结合建模、行为分析、基因操作等方法，探讨对视觉特征提取、处理、整合的机制。主要回答：小尺度物体运动敏感的小视场系统的神经基础是什么？小视场系统和对背景运动敏感的大视场系统是否独立？位置系统编码特性是什么？研究结果期待揭示神经元和环路机制，为实时检测与跟踪的人工智能提供神经计算原理依据。

脑科学和类脑智能研究的重要趋势是“了解人类智力进而开发智能机器”（美国MIT IQ计划，2018）。我们正面临人工智能革命新一轮的“生存竟争”风暴。今天神经科学重要进展之一，是绘制了从果蝇视网膜到髓质这一与初级运动检测器环路对应的神经网络结构，证明果蝇和小鼠的运动视觉计算惊人相似（Nature Neurocience, 2015）。然而，对初级运动检测器突触后的高级视叶区仍知之甚少。.本课题聚焦果蝇复眼运动视觉系统，采用基因操作阻断特定神经元、果蝇定量行为分析、构建神经环路计算模型、以及大规模数值仿真实验等方法相结合的研究途径，揭示视觉特征检测的神经元和环路机制，提出复眼视觉系统启发的目标实时检测算法。.本课题的研究进展包括，初步鉴定了果蝇复眼视觉系统中参与高阶运动检测、或者离散视觉线索组成的复合图形检测的数种视觉投射神经元；提出了基于EMD局域视觉运动估测的图形‒背景分离的计算神经理论模型；抽提了基于生理约束的迫近物体检测的算法模型，并用虚拟机器人验证了该算法具有实时应用潜力。本课题的研究进展，不仅对果蝇初级运动检测器突触后的高级视觉区的特征提取、处理、和整合的神经机制的理解进了一步，而且提出了神经计算原理启发的、具有实时应用潜力的类脑人工智能算法。