Internet-of-Things (IoT) is beginning to have a wide impact across various sectors of the economy.  Traditionally, data are collected in an IoT system from various sensors and then transmitted to the edge devices and the cloud where machine learning models are trained and used to provide inference. This process, unfortunately, consumes too much energy and bandwidth.  Hence, it is not sustainable.  Recognizing this, researchers have begun to push intelligence downwards from the cloud to edge devices.  Since many edge devices are energy-constrained, this means new approaches to energy-efficient inference need to be developed.  This was the goal of this research project.

The project resulted in several new methodologies in this domain.  Some examples are: energy/memory/latency-efficient neural network synthesis, incremental training of neural networks, introducing hidden layers in the control gates of long short-term memories (LSTMs) to make them more accurate and energy/latency-efficient, augmenting feature space with semantic space data for improving classification accuracy, understanding the physical/mental/emotional states of an individual through wearable medical sensors (WMSs), using WMSs for diabetes diagnosis (both Type I and II) on an edge device, convolutional autoencoder based transfer learning, use of random projections for energy efficiency, using genetic programming for feature computation, etc. These methodologies significantly extend the state-of-the-art in the respective research areas.

Eight PhD students (four females) and nine undergraduate students (two females) were trained in this exciting area of research.  Their work resulted in several journal articles.  The results were disseminated to the wider public through various invited talks. Some of the results have been licensed to a company for commercialization.  The students also did technology transfer through their summer internships at various companies. 

Last Modified: 10/19/2020
Modified by: Niraj K Jha