Computational modelling of transformation invariant representations and information maintenance in spiking neural networks with ranges of axonal condu

具有轴突传导范围的尖峰神经网络中变换不变表示和信息维护的计算模型

• 批准号: 1789468
• 金额: --
• 依托单位: University of Oxford
• 依托单位国家: 英国
• 项目类别: Studentship
• 财政年份: 2016
• 资助国家: 英国
• 起止时间: 2016 至 无数据
• 项目状态: 已结题
• 来源: https://gtr.ukri.org/projects?ref=studentship-1789468
• 关键词: Computational; modelling; transformation; invariant; representations

The application of computational visual neuroscience to the modelling of higher-level brain function and practical robotics has been limited by the absence of stereoscopic primate models of high-level, three-dimensional vision. It is only recently that neurophysiological characterisations of three-dimensional object shape representation in the macaque inferior temporal cortex (IT) have emerged, providing results for computational models to replicate. The role of IT's visual representation in the fundamental functions of visual attention (Chelazzi et al., 1993), object recognition (Hung et al., 2005) and object interactions (Gallese et al., 1994), makes its replication as a modular interface between reality (both physical and virtual) and higher-level modelling, of significant interest to artificial intelligence and of critical importance to future computational neuroscience research.This research proposal plans to replicate the IT representation of three-dimensional object shape found by Yamane et al. (2008), within the hierarchical VisNet model of primate vision. This progresses the research of Eguchi et al. (2015) of the OCTNAI in replicating characterisations of two-dimensional object shape representation in V4 and posterior IT (TEO), within the VisNet model. The adaptive stimuli technique and analysis provided by Yamane et al. will allow for a direct comparison of the model with neurophysiological results. This work will also provide the basis for modelling recent characterisations of more advanced representations of object shape in IT. The functionality of the initial model's representations for higher-level processing will be tested by modelling the theorised role of the anterior intraparietal cortex (AIP) in mapping IT shape representation to motor commands for grasping objects. This will incorporate the research of the OCTNAI in modelling head-centred (Mender and Stringer, 2014) and hand-centred (Galeazzi et al., 2015) egocentric reference frames with the aim of producing a practical model of visually guided reaching and grasping.As the central area of investigation, the proposal begins by outlining characterisations of three-dimensional object shape representation in IT and describes hypotheses of how such representations may self-organise within a stereoscopic version of VisNet. This is followed by an outline of my current internship research, which has aimed to produce a biologically accurate model of stereoscopic disparity in V1, to act as a model of binocular combination for input into higher-level VisNet experiments. In particular, it focuses on the extensions I have made to the Topographica model of stereoscopic vision in V1, in order to improve the biological accuracy of its representation of disparity in line with recent neurophysiological results. Characterisations of disparity tuning in V2 and V4 are also discussed, which will act as intermediate benchmarks for the model. The proposal specifies experiments to test the theorised role of AIP in reaching and grasping. Finally, details of my suitability for undertaking this research are outlined.

由于缺乏高级三维视觉的立体灵长类动物模型，计算视觉神经科学在高级大脑功能和实用机器人建模中的应用受到限制。直到最近，猕猴下颞叶皮层（IT）中三维物体形状表征的神经生理学特征才出现，为计算模型的复制提供了结果。 IT 的视觉表示在视觉注意（Chelazzi 等人，1993）、物体识别（Hung 等人，2005）和物体交互（Gallese 等人，1994）等基本功能中的作用，使其复制为现实（物理和虚拟）和更高层次建模之间的模块化接口，对人工智能具有重大意义，对未来的计算神经科学研究也至关重要。该研究计划计划 复制 Yamane 等人发现的三维物体形状的 IT 表示。 (2008)，灵长类视觉的分层 VisNet 模型。这推进了 Eguchi 等人的研究。 (2015) OCTNAI 在 VisNet 模型中复制 V4 和后验 IT (TEO) 中二维对象形状表示的特征。 Yamane 等人提供的自适应刺激技术和分析。将允许将模型与神经生理学结果进行直接比较。这项工作还将为 IT 中更高级的对象形状表示的最新特征建模提供基础。通过对前顶内皮层 (AIP) 在将 IT 形状表征映射到抓取物体的运动命令方面的理论作用进行建模，将测试初始模型的高级处理表征的功能。这将结合 OCTNAI 在以头为中心（Mender 和 Stringer，2014）和以手为中心（Galeazzi 等人，2015）的自我中心参考系建模方面的研究，旨在产生视觉引导的触及和抓取的实用模型。作为研究的中心领域，该提案首先概述了 IT 中三维对象形状表示的特征，并描述了如何实现这一点的假设。 表示可以在 VisNet 的立体版本中自组织。接下来是我当前实习研究的概述，该研究的目的是在 V1 中生成生物学上精确的立体视差模型，作为双目组合模型输入到更高级别的 VisNet 实验中。特别是，它重点关注我对 V1 中的立体视觉地形模型所做的扩展，以便根据最近的神经生理学结果提高其视差表示的生物学准确性。还讨论了 V2 和 V4 中视差调整的特征，这将作为模型的中间基准。该提案指定了实验来测试 AIP 在到达和抓取方面的理论作用。最后，概述了我是否适合进行这项研究的详细信息。