Neural networks (NNs) have become the workhorse of the Artificial Intelligence (AI) industry.  A decade ago, the trend was to design large NNs trained on big data.  However, soon it was realized that edge-friendly NNs could have much broader applications.  These NNs have to be highly energy-, memory- and latency-efficient since many edge devices are battery-operated. Such devices include smartphones and smartwatches.  This project was focused on the design and applications of such NNs.

Many strong pillars are needed to support edge-friendly NNs.  The main one being a tool that synthesizes energy/memory/latency-efficient NNs from given datasets.  We have developed a tool, called SCANN, that results in NNs that are, on an average, two orders of magnitude smaller than traditional fully-connected NNs, yet at the same time boost the NN accuracy.  SCANN introduced a grow-and-prune synthesis paradigm modeled after the human brain.  Starting from a seed architecture, akin to a baby brain, it adds neurons and connections to the NN to significantly improve its accuracy, ending in an NN that is akin to a toddler brain.  Finally, SCANN specializes this NN into an adult-brain-like architecture that performs very accurately on the given application, but is orders of magnitude smaller.  We extended SCANN to a tool called CURIOUS to fully exploit the grow-and-prune synthesis paradigm.

Edge applications typically do not have large training datasets.  However, NNs are very data-hungry.  Thus, another important pillar is a synthetic data generator that can sample unlimited synthetic data instances from the same probability distribution as the real dataset.  We have developed one such tool, called TUTOR.  It provides a significant boost to accuracy of NNs.  In medical applications, it also ensures privacy when real data cannot be shared.  Instead, synthetic data can be shared without sacrificing on the accuracy of the models trained with such data.

A third pillar is finding out which data instances are incorrectly labeled.  In supervised machine learning, it is extremely important that the labels be correct.  Otherwise, prediction accuracy suffers.  We developed a tool, called CTRL, that automatically determines which labels are wrong.  Removing the data instances with wrong labels boosts the accuracy of NNs significantly.   CTRL is applicable to both tabular and image-based datasets. Thus, it is widely applicable across diverse Internet-of-Things (IoT) applications.  It is an important preprocessing step before NN training commences.

Real-world datasets often have missing values. Since the dataset sizes are often small, it is not simply possible to delete all data instances that have any missing values.  Thus, imputing these values becomes important.  We have developed a tool, called DINI, that imputes missing values and obtains state-of-the-art results in terms of NN accuracy.

Edge-friendly NNs make new smart healthcare applications possible.  We have developed a framework, called DOCTOR, that produces a multi-headed NN.  The body of the NN can be trained with wearable sensor data that can be obtained from sensors found on the smartwatch or smartphone and the different heads can detect different diseases.  We have shown how DOCTOR can detect Covid-19, Type I/II diabetes, and mental health disorders (e.g., depression, bipolar disorder, and schizophrenia) through the multiple heads.  This multi-headed NN is highly energy/memory/latency-efficient and can reside on a smartwatch.

Convolutional neural networks (CNNs) are often used in image-based applications. These are often very large models and thus their inference latency is often not compatible with latency-sensitive applications.  We have developed a dynamic inference method that skips many layers of the CNN depending on the image it encounters at inference time, without giving up on prediction accuracy. 

Four PhD students 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. The students also did technology transfer through their summer internships at various companies.  The frameworks were used in various course projects.  Several tools/frameworks developed in this project are also being commercialized by a startup company.

Last Modified: 10/15/2023
Modified by: Niraj K Jha