Adaptive Learning Methods for Deeply Embedded Devices

点击翻译按钮获取中文摘要

Smartphones have changed our everyday life, but an even greater revolution is at our door. The future will be composed of tiny computing devices deeply embedded in our environment (e.g., houses, vehicles, clothes, roads, workplaces), which will sense and process data, and communicate with other devices and the Cloud. The advances in electronics, sensors, networking, and battery technologies will make it possible to build such devices, while the software technologies supporting this new class of computers is far from ready. Therefore, our research will explore avenues to improve software approaches relevant to this computing platform, by proposing efficient methods for processing the sensed data on the devices while dealing with the limited resources that they provide. Three aspects will be developed in the research. First, methods grounded in artificial intelligence will be investigated to allow an efficient processing of the data sensed by the devices. Let us assume a given object of interest, for which each device has a different view. The approach followed is to process the sensed data locally, on the devices and according to their respective view, before reaching a decision regarding the class of object at the network level. That would involve starting with a general recognition model, which is produced from large databases of object observations, and then specializing the model in an online fashion, according to the context and environment in which each device is operating. As a second aspect, we will develop methods to provide self-managing capabilities to the devices. This should lead to an improvement in the performance of the devices and increase their energy autonomy, all this with little or no human intervention. Our third objective is to establish methodologies to design the software system used on these devices in accordance with their capabilities and duties. For that, we will develop techniques that will explore the different possible trade-offs in the design of the systems. Indeed, with these devices, an increase in processing so as to achieve better performance generally leads to the consumption of more resources, possibly going beyond what is available or significantly reducing their energy autonomy. A better understanding of the impact associated with better sensing performance should allow the system designer to make a more informed decision. This research is necessary to support the exploitation of deeply embedded devices, required for the so-called ambient intelligence paradigm to become a reality. In an ambient intelligence world, the embedded devices are helping people to carry out their everyday activities in an unobtrusive and natural way. These miniature devices will be well integrated in the environment, such that they will disappear from our surroundings, the only visible element being the user interface. The impact of this new computing paradigm on our everyday life will be tremendous in terms of productivity (e.g. automation, resources management, quality of life), safety (e.g. transportation, public security), and health (e.g., disease detection, first response). The technologies developed in this programme should provide Canada with a competitive edge in this domain, creating intellectual property and expertise for developing software that is more robust, capable of adaptation and self-management, which are key qualities to allow computing to move out of the office and enter the real-world.