CIF: Small: Task-Cognizant Sparse Sensing for Inference

随着大数据应用程序的规模和数量不断增长，仅处理对特定推理任务有信息的测量结果至关重要，以便限制所需的传感成本以及存储或传输数据的相关成本。为此，该项目为随机过程开发了一种新的压缩感知范式，其中信号压缩和重构策略被设计为任务识别，依赖于有用的信号统计而不是原始的随机信号。这项研究导致主要的传感节能，并可以受益于大量的节能传感应用，如位置感知服务，天气监测，频谱监测和射电天文学。该项目的主要目标是通过提供创新的传感方法，为感兴趣的推理任务量身定制，以显着降低传感成本以及相关的存储和通信要求。与现有的压缩采样不同，所提出的研究利用信号统计信息的固有结构，以便能够从稀疏采样数据中进行可靠的推断，而不是仅仅依赖于信号稀疏度来实现压缩。因此，一种新的方法来稀疏感测引入随机过程，这揭示了压缩的基本限制相对于固有的基本统计结构的自由度，即使在没有信号稀疏。提出了有效的推理技术和确定性压缩采样器设计，以在实现期望的推理质量的同时影响感测成本的主要节省。这些基本的结果开辟了机会，有效地处理数据密集型传感应用程序中，从随机过程的推理是最重要的。

As big data applications continue to grow in size and number, it is of crucial importance to deal only with measurements that are informative for a specific inference task in order to limit the required sensing cost, as well as the related costs of storing or communicating the data. To this end, this project develops a new paradigm of compressive sensing for random processes, where the signal compression and reconstruction strategies are designed to be task-cognizant, hinging on useful signal statistics rather than the original random signals. This research leads to major sensing energy savings, and can benefit a plethora of energy-efficient sensing applications such as location-aware services, weather monitoring, spectrum monitoring, and radio astronomy.The main goal of this project is to significantly reduce the cost of sensing as well as the related storage and communication requirements by offering innovative sensing approaches tailored to the inference task of interest. Different from existing compressive sampling, the proposed research leverages the inherent structures of signal statistical information in order to enable reliable inference from sparsely sampled data, rather than solely relying on signal sparsity to enable compression. As such, a new approach to sparse sensing is introduced for random processes, which reveals the fundamental limits of compression in relation to the degrees of freedom inherent to the underlying statistical structure, even in the absence of signal sparsity. Efficient inference techniques and deterministic compressive sampler designs are put forth to affect major savings in the sensing costs while achieving the desired inference quality. These basic results open up opportunities for efficient handling of data-intensive sensing applications in which inference from random processes is of foremost importance.