Personalized Dynamic Attention Tracking

拟议的研究计划旨在通过开发、实现和验证视觉注意的计算模型来进一步提高视觉注意跟踪系统的能力，该计算模型估计或预测视觉注意随时间的分配(即，人们在看的地方)。注意力跟踪有许多有用的应用，例如在开发智能交互显示、在零售环境中确定购物者的偏好和偏见，或者为标志或网页创建有效的视觉布局。这项研究的主要焦点将是注意力模型的“个性化”。最好的视觉注意计算模型是基于深度神经网络的。这些神经网络是在从大量观众那里获得的眼球运动的广泛数据库上进行训练的。大量的训练样本意味着可以在不过度拟合的情况下训练大容量网络，从而可以很好地预测人们对新的、前所未有的图像的注意力。然而，注意力模型的许多重要应用领域，如智能显示设备，都是针对个人用户的。例如，手机通常是一个人使用的，汽车是由少数人驾驶的。在这些应用中，注意力模型很好地适应了大量人群的平均行为，可能不能很好地预测个人的注意力分配。因此，我们的研究将集中于开发能够使大量人口的“普通人”注意模型适应于在人群的子集(例如男性/女性或老年人/年轻人或个体)上表现良好的模型的技术。我们还将调查注意偏差的检测，这些偏差可能是个体群体中常见的。为了实现个性化注意模型的目标，我们的PI研究团队、两名博士生、一名硕士学生和本科生暑期实习生将利用最新的机器学习技术，如领域适应，利用稀疏的个性化数据来修改大群体模型，以更好地适用于较小的、个性化的群体。拥有个性化的注意力跟踪可以让显示器受益，比如汽车仪表盘或手机上的显示器。在你的车里，仪表盘会识别你并调整仪表盘的显示，以提供重要的信息来有效地吸引你的注意力，同时以不那么分散注意力的方式呈现其他信息。另一个影响领域将是增强无处不在的技术，人们在日常生活中与之互动的次数越来越多。例如，在你最喜欢的零售店，根据你的注意力集中模式，商店将能够预测你的注意力偏见，同时保持你的隐私，允许它基于这些偏见提供服务(展示有用的信息、提供折扣和忠诚度奖励)。个性化的注意力跟踪有望使人与机器之间的联系更加自然。

The proposed research program aims to further the capabilities of visual attention tracking systems by developing, implementing and validating computational models of visual attention that estimate or predict the allocation of visual attention over time (i.e. where people are looking). Attention tracking has many useful applications, such as in developing intelligent interactive displays, determining shoppers' preferences and biases in a retail setting, or creating effective visual layouts for signs or webpages. The primary focus of the research will be on the "personalization" of attention models. The best computational models of visual attention are based on deep neural networks. These neural networks are trained on extensive databases of eye movements obtained from large numbers of viewers. The large number of training examples means that high capacity networks can be trained without over-fitting, leading to good prediction of attention on new, never-seen-before, images. However, many of the important application areas of attention models, such as intelligent display devices, are focused on individual users. For example, cellphones are typically used by a single person and cars are driven by a small number of people. In these applications, an attention model that is tuned well to the average behaviour of large populations, may not work well in predicting attention allocations of individuals. Therefore, our research will focus on developing techniques that can adapt the large population "average person" attention models to ones that perform well on subsets of the population (e.g. male/female or old/young or individual persons). We will also investigate detection of attentional biases that may be common to groups of individuals. To achieve our objective of personalized attention models, our research team of the PI, two PhD students, a Masters student, and undergraduate summer interns will employ recent machine learning techniques such as domain adaptation to make use of sparse individualized data to modify large population models to work well for smaller, personalized, groups. Having personalized attention tracking can benefit displays such as those found in automobile dashboards or cellphones. In your car, the dashboard will recognize you and adjust the dashboard display to provide important information to effectively capture your attention, while presenting other information in less distracting fashion. Another area of impact will be in enhancing ubiquitous technology that people interact with more and more in every day life. For example, in your favorite retail outlet, based on your patterns of attentional focus, the store will be able to predict your attentional biases while maintaining your privacy, allowing it to provide services (presenting helpful information displays, offering discounts and loyalty rewards) based on these biases. Personalized attention tracking promises to make the connection between human and machine more natural.