Computational modelling visual perception in a biologically realistic neural network: Developing rich, hierarchical representations of visual scenes

许多计算神经网络已经解决了视觉对象识别的任务；然而，仅靠这一点是不足以理解视觉世界的。还需要形成场景内的特征和对象的分层表示，对它们彼此之间的语义关系进行编码(例如，“车轮”是“汽车”的一部分)。这就是心理学中的约束问题：当存在多个对象时，我们如何知道哪些特征是哪些对象的一部分？最近提出了一种潜在的结合问题的解决方案，以模型网络的形式结合了四个关键的生物学特征，这些特征使其有别于以前的计算机视觉网络：(I)自上而下和侧向连接，(Ii)尖峰动力学，(Iii)随机分布的轴突延迟，(Iv)依赖于尖峰时间的可塑性。这种网络有可能以一组神经元的形式对刺激进行编码，这些神经元以时间上精确的序列(多时神经组)发射，并且已经被证明只有在低级别特征表征驱动高级特征表征时才能产生发射的神经元。然而，到目前为止，只有非常有限的刺激集用于培训，并采用了相对表面的分析。我建议在比以前使用的更生态有效的刺激集上训练这个网络，包括3D对象和多对象场景的图像，并对通过网络的各个层发展的表示进行深入分析，这些表示由单个神经元和多时相组编码。这样的项目将朝着创建一个生物逼真的计算机网络迈出一步，这种计算机网络可以真正意义上的视觉世界。这将有很大的实用价值，并提供对灵长类视觉系统如何发展视觉特征正确绑定的分层表征的洞察：这是视觉心理学中一个长期存在的关键问题。

The task of visual object recognition has been solved by many computational neural networks; however, this alone is insufficient for making sense of the visual world. It is also necessary to form hierarchical representations of the features and objects within a scene, encoding their semantic relations to one another (e.g. "wheel" is part of "car"). This is the binding problem in psychology: when multiple objects are present, how do we know which features are part of which objects? A potential solution to the binding problem has recently been proposed, in the form of a model network which incorporates four key biological features that set it apart from previous computer vision networks: (i) top-down and lateral connections, (ii) spiking dynamics, (iii) randomly distributed axonal delays, and (iv) spike-timing dependent plasticity. This network has the potential to encode stimuli in the form of groups of neurons which fire in a temporally precise sequence (polychronous neural groups), and has been shown to develop neurons that fire only if a low-level feature representation is driving a high-level feature representation. So far, however, only very limited stimulus sets have been used for training and relatively surface-level analyses employed. I propose to train this network on a more ecologically valid stimulus set than that previously used, including images of 3D objects and multi-object scenes, and carry out an in-depth analysis of the representations that develop through the layers of the network, encoded both by individual neurons and polychronous groups. Such a project would take steps towards the creation of a biologically- realistic computer network that can truly make sense of the visual world. This would be of much practical use, and provide insight into how the primate visual system develops hierarchical representations in which visual features are correctly bound: a long-standing and crucial problem in visual psychology.