Efficient neural network synthesis and its applications to smart healthcare

Dec 9, 2021, 2:00 pm2:00 pm



Event Description

Deep neural networks (DNNs) have become the driving force behind recent artificial intelligence research. An important problem with implementing a neural network is the design of its architecture. Another issue is that modern neural networks often contain millions of parameters, whereas many applications require small inference models due to imposed resource constraints, such as energy constraints in battery-operated devices. However, efforts to migrate DNNs to such devices typically entail a significant loss of classification accuracy. To address these challenges, we first propose a two-step neural network synthesis methodology, called DR+SCANN, that combines two complementary approaches to design compact and accurate DNNs. At the core of our framework is the SCANN methodology that uses three basic architecture-changing operations, namely connection growth, neuron growth, and connection pruning, to synthesize efficient feed-forward architectures with arbitrary structure. These neural networks are not limited to the multilayer perceptron structure. Furthermore, we propose an approach to synthesize accurate DNN models with limited available data, using synthetic data drawn from the same probability distribution as the real data.

We also look into the application of the efficient DNNs to smart healthcare, e.g., rapid diagnosis of the SARS-CoV-2 virus and the resultant COVID-19 disease. The current testing regime based on Reverse Transcription-Polymerase Chain Reaction for SARS-CoV-2 has been unable to keep up with testing demands. Hence, there is a need for an alternative approach for repeated large-scale testing of SARS-CoV-2/COVID-19. We propose a framework called CovidDeep that combines efficient DNNs with commercially available wearable medical sensors (WMSs) for this purpose. We collected data from individuals spanning three cohorts: healthy, asymptomatic (to detect the virus), and symptomatic (to detect the disease). We trained DNNs on various subsets of the features automatically extracted from six WMS and questionnaire categories to perform ablation studies to determine which subsets are most efficacious in terms of test accuracy for a three-way classification. The highest test accuracy obtained was 98.1%. The resultant DNNs are embedded in a smartphone application, which has the added benefit of preserving patient privacy.