Towards Cognizant Sensors: Making sense of data through physics

分布式传感器网络，如可穿戴设备和物联网，产生越来越多的数据，这些数据被本地或远程收集和处理，以使用复杂的机器学习算法生成上下文。发展传感系统的下一步是在传感器层面整合传感和认知能力，以便在数据到达时检测到数据中的模式，而无需来回传输原始数据。该研究项目建立在使用非线性材料和设备响应作为非常规学习平台的最新进展的基础上，开发传感系统，不仅可以测量目标参数，还可以同时识别模式并提供认知信息；即，认知传感器。我们的方法是基于一种被称为“水库计算”的技术，在这种技术中，输入被应用到一个由相互连接的非线性神经元（即水库）组成的固定网络中，该网络以一种衰减的方式保留有关过去事件的信息。储层的功能是将输入数据非线性地映射到高维空间，使目标事件是可分离的。然后通过一个简单的接口读取神经元的输出，并使用可训练的算法自适应地组合。由于储层在整个过程中保持不变，因此储层计算是物理机器学习的合适平台。这种基于非线性光子、电子和流体神经元的系统的例子已经提出并证明了。

Distributed sensor networks, such as wearables and Internet-of-Things, produce growing amounts of data that is collected and processed locally or remotely to generate context using sophisticated machine learning algorithms. The next step in developing sensing systems is to integrate the sensing and cognition abilities at the sensor level so that patterns in the data are detected as it arrives without the need for back and forth transmission of raw data. This research program builds upon the latest advances in using nonlinear material and device responses as unconventional learning platforms to develop sensing systems that not only measure the target parameters but simultaneously recognize patterns and provide cognitive information; i.e., Cognizant Sensors. Our approach is based on a technique known as Reservoir Computing, where the input is applied to a fixed network of interconnected nonlinear neurons (i.e., the reservoir) that retain, in a decaying manner, information about the past events. The function of the reservoir is to map the input data onto a high-dimensional space nonlinearly such that the target events are separable. The outputs of neurons are then read by a simple interface and adaptively combined using a trainable algorithm. As the reservoir remains unchanged throughout the process, reservoir computing is a suitable platform for physical machine learning. Examples of such systems based on nonlinear photonic, electronic, and fluidic neurons have been proposed and demonstrated. We will rely on our proven track record in the fields of micro- and nano-systems, statistical signal processing, and study of various nonlinear phenomena to conduct the fundamental research required to develop physical reservoirs and apply them to solve three selected challenges in the field. Specifically, we will develop: (i) A Cognizant Accelerometer for the condition monitoring of machinery; (ii) A Cognizant Gas Sensor to address the cross-sensitivity challenge when trying to measure a particular gas concentration in a gas mixture; and (iii) A Cognizant Textile that can detect human activity once worn as part of a garment. The core knowledge and techniques developed through this program apply to other areas at micro- or macro-scales. The trainees involved in this program will collaborate closely with each other and will attain skillsets that expand beyond the framework of their individual projects. The program is expected to result in new solutions and design guidelines for Cognizant Sensors through follow up projects and will instigate future collaborations with academic and industrial partners. Such partnerships provide internship and hands-on training opportunities to the trainees, broadening their technical, professional, and entrepreneurial horizons.