Machine Learning with Uncertainty for Monitoring Moving Objects and People

序列数据在自然界中非常常见，出现在无数的数据集中，如股票市场、天气、生物医学、目标跟踪等。序列通常由多个随时间一起变化的相关信息通道组成。最先进的序列数据深度学习模型使用标量来表示数据，而不是随机变量。这些基于标量的深度学习模型需要大量的数据来训练，当在一个领域的数据上训练然后应用到另一个领域时表现不佳，并且对罕见事件的建模很差。序列数据的传统统计模型不是基于学习的，通常依赖于固定的参数，这些参数应该根据不同的环境或不同的情况进行调整。贝叶斯概率模型允许设计可解释的系统，处理小数据和量化估计和预测中的不确定性。新的概率深度学习框架为这一领域的发展奠定了基础。利用贝叶斯概率模型和深度学习模型的研究最近才刚刚开始。在这个项目中，我们计划为开发时间序列和序列数据的概率深度学习模型做出重大贡献。这一领域的初步工作产生了强大、通用但非常复杂的模型。我们计划开发更简单的模型和质量度量，并使它们更容易训练，以及在实际应用中实时实现模型。我们将在至少两个实际应用中测试开发的模型构建方法：监测人们的活动和意图，重点是监测老年人，以及分类和跟踪多架无人机并推断其意图。对于数据收集，我们计划在两个应用程序中使用相同的传感器，并使用非常相似的算法框架。识别意图非常重要，因为这使机器学习更接近于人类观察他人行为和移动物体的方式。这项研究将通过整合贝叶斯概率学习、时间序列、深度学习和不确定性量化等多个概念，推动机器学习领域的发展。如果开发的模型构建框架优于目前的技术水平，对加拿大和世界的好处将是巨大的，对加拿大的额外好处是在一个非常受欢迎的领域培养10名学生，以及通过更准确地监测运动物体带来的风险来推进老年人护理和改善公共安全的新应用。

Sequence data is very common in the natural world, emerging in myriad datasets such as stock markets, weather, biomedical, target tracking and so on. Often the sequences are composed of multiple channels of correlated information that change together in time. State of the art deep learning models of sequence data represent data using scalars, rather than random variables. These scalar-based deep learning models require large amounts of data to train, perform poorly when trained on data from one domain and then applied to a different domain, and model rare events poorly. Traditional statistical models for sequence data are not learning-based and very often rely on fixed parameters that should be adjusted for different environments or different situations. Bayesian probabilistic models allow for designing interpretable systems, dealing with small data and quantifying uncertainties in estimations and predictions. New probabilistic deep learning frameworks have primed this field for advancement. Research on taking advantage of both Bayesian probabilistic and deep learning models has just started recently. In this program, we plan to significantly contribute towards developing probabilistic deep learning models for time series and sequence data. Initial work in this field resulted in powerful, versatile but very complex models. We plan to develop simpler models and quality metrics, and to make them easier to train as well as to implement the model in real-time in practical applications. We will test the developed model building approaches on at least two practical applications: monitoring the activities and intents of people with the focus on monitoring elderly people, as well as classifying and tracking multiple UAVs and inferring their intents. For data collection, we plan to use the same sensors for both applications, and a very similar algorithmic framework. Recognizing intents is extremely important because this brings machine learning closer to the way that humans observe the behaviour of other humans and moving objects. This research will move the field of machine learning by integrating multiple concepts including Bayesian probabilistic learning, time series, deep learning, and uncertainty quantification. The benefits to Canada and the world would be tremendous if the developed model building framework outperforms the state of the art, and additional benefits to Canada are the training of ten students in a highly sought after field, and new applications advancing the care of elderly people and improving public safety by more accurately monitoring the risk posed by objects in motion.