Attention and Selection Mechanisms in Neural Networks and Computer Vision

研究内容包括神经信息处理系统中信息的注意、选择、门控和路由机制，包括视觉问题的顺序解决方案。循环深度学习模型构成了展示所提议的工作语料库的价值和效用的主要结构。虽然深度学习模型在广泛的问题上取得了相当大的成功，但仍然需要更好地描述它们的行为，并在产生解决方案的方向上进行创新，这些解决方案在实现其目标时既节俭又有效。为此，本建议包含了一个全面的理论和实验工作体系，以实现这些目标。这是通过两个不同但互补的支柱作为基础来实现的：首先，我们建议详细检查通过神经层次控制信息流的循环门控机制。这包括添加剂。乘法和其他门控机制，它们与增益控制和归一化策略的关系，以及与常见前馈网络中出现的规范机制的联系。我的研究小组已经展示了这种深度神经网络能力控制机制的巨大潜力。其次，我们需要考虑解决包括相关模式或突出区域的顺序或时间识别在内的问题。视觉本质上是连续的，最明显的例子是人类的头部和眼睛运动。同样重要的是大脑处理过程的动态特性，包括注意力的作用。迄今为止，计算机视觉中的许多问题都是以一种一次性对图像进行逐像素分配的方式进行操作的。第二个支柱包括仔细检查空间中的注意选择，以及信息处理的循环机制，以促进视觉处理。这项研究的成果将包括凝视模式的顺序模型，以顺序的方式解决视觉问题，焦点处理和循环门控之间的关系，以及空间采样的人类和机器视觉解决方案的比较。这项工作的每个方面都从建立重要性的基本原则作为基础开始，同时检查循环网络的实际适用性，并绘制与注意机制的联系。作为一个整体，拟议的研究将从架构的角度和在增强我们对神经网络信息处理能力的理解的层面上为深度学习提供基础贡献。此外，该研究将现状扩展到机器视觉问题的更广泛的顺序或时间解决方案，作为问题解决的替代范例，并以一种适合主动视觉系统的形式。

The proposed research is comprised of mechanisms for attention, selection, gating and routing of information in neural information processing systems including sequential solutions to visual problems. Recurrent deep learning models form the primary construct for demonstrating the value and utility of the proposed corpus of work. While deep learning models have shown considerable success on a wide range of problems, there remains a need for better characterization of their behaviour, and innovation in directions that produce solutions that are both more parsimonious and effective in addressing their objectives. To this end, this proposal encapsulates a comprehensive body of theoretical and experimental work to meet these objectives. This is achieved through two distinct but complementary pillars as a foundation: First, we propose to examine in detail recurrent gating mechanisms that control flow of information through neural hierarchies. This includes additive. multiplicative and other gating mechanisms, their relation to gain control and normalization strategies, and connections to canonical mechanisms that appear in common feed-forward networks. My research group has shown significant potential for such control mechanisms for the capabilities of deep neural networks. Second, we address the need to consider solutions to problems that include sequential or temporal identification of relevant patterns or salient regions. Vision is inherently sequential with the most palpable example of this in humans involving head and eye movements. Equally important is the dynamic nature of processing within the brain including the role of attention. To date, many problems in computer vision operate in a manner that produces pixel-wise assignments to an image in one pass. This second pillar involves careful examination of attentive selection in space, and recurrent mechanisms for information processing to facilitate processing for vision. Products of this investigations will include sequential models of gaze patterns, solution to vision problems in a sequential manner, relationships between focal processing and recurrent gating and comparison of human an machine vision solutions to spatial sampling. Each facet of this work begins with establishing basic principles of importance as a foundation while examining practical applicability in recurrent networks and also drawing connections to attention mechanisms. As a whole, the proposed research stands to provide fundamental contributions to deep learning from an architectural perspective and at the level of augmenting our understanding of information processing capabilities of neural networks. Moreover, the research extends the status quo into a broader set of sequential or temporal solutions to machine vision problems, as an alternative paradigm for problem solving and in a form amenable to active vision systems.