DiagnoseNET – NVIDIA Jetson Developer Challenge
Articles Blog

DiagnoseNET – NVIDIA Jetson Developer Challenge

August 9, 2019

The value of data in research health is the possibility of new answers for treatment, prevention and comprehension of disease and quality of life. A common challenge in current medical diagnosis is that physicians have access to massive amount of data on patients, but have the short time to analyse all these data. One limitation is that hospitals didn’t have robust computational resources for training large deep neural networks, an alternative is using the AI public cloud market. Therefore the patient’s data need use anonymizing data to reducing disclosure risk when making data available for research, this procedure could lose the truthfulness at the record level and cause wrong mining results. We are building a framework called DiagnoseNET for training full Deep Neural Networks on Mini-Cluster Jetson TX2 Inside the Hospitals. This cluster can be used as a learning center with minimal infrastructure requirements, low power consumption and brings the opportunity to use a dataset with more patient’s feature DiagnoseNET provides three high-level features: a framework to build full Deep Neural Networks (DNN) workflow, a distributed processing training DNN on Mini-Cluster Jetson TX2, and an energy-monitoring tool for workload characterization. We are building a Mini-Cluster Jetson TX2, starting with 8 Jetson boards and we have planned to work with 12, 24 boards to evaluate their performance and power consumption. We started with the implementation of the distributed processing at stage two from the full DNN workflow, that we called DiagnoseNET unsupervised embedding. To exploit the parallel and distributed processing we selected the data parallelism, where we split the binary patient phenotype representation in several mini-batch and assigned each group to one jetson board. The first technique implemented is the synchronous training, were all replicas task read the same values for the current parameters, compute gradients in parallel, and theirs apply together. DiagnoseNet as a framework to build full Deep Neural Networks workflow is divided into three stages: DiagnoseNet Data-Mining: to drive the Binary Patient Phenotype Representation DiagnoseNet Unsupervised Embedding: to get a new encoded space of patient phenotype. DiagnoseNet Supervised Learning: for classification and prediction of medical Inpatient states at Intensive Care Unit

Leave a Reply

Your email address will not be published. Required fields are marked *