Adaptive Resolution Imaging Sonar: Initial Exploration at the Shiawassee National Wildlife Refuge
The ARC Scientific Computing and Research Consulting Services supported a team of masters students from the School of Environment and Sustainability (SEAS) in developing a custom solution to conduct their research.
The ARC Consulting Services team provided support in grant writing, data pipeline development, machine learning training, and evaluation of a custom solution to identify fish in a sonar camera stream. Jetstream was used to store and process specialized data. The Consulting team also built on the ResNet-50 neural network architecture to train and run a neural network to locate and determine directionality of fish in the camera feed.
View Long Description here
DDPG Training: Optimal Parameter Design in Circuits
ARC worked with a PhD student team to design a Deterministic Deep Policy Gradient agent to determine the optimal parameters for circuits to achieve the highest power efficiency. The DDPG training environment was a simulation block of a circuit model, which calculates the power efficiency given circuit parameters. This policy was implemented in Matlab, using a training reward as a function of power efficiency.
Automating the Pipeline from Glomerular Detection to Morphometric Measures
ARC is currently collaborating with a research team from CSCAR, along with the med school, to automate an approach to measure glomerular morphometry using neural networks along with segmentation and masking. This will allow for assessment of a large number of glomerular and other features from nephrectomy sections and kidney biopsies. We are using a convolutional neural network for deep learning classification of glomerular structures as well as segmentation for boundary detection and masking of glomerular areas.
View Long Description here
Alternative models for measuring biomass in practice
ARC consultants worked with a Master’s student team from the School for Environment and Sustainability (SEAS) comparing five modeling techniques — ordinary least squares (OLS) regression, linear fixed effects (FE), power law, random forest (RF), and support vector regression (SVR) — for estimating biomass from LiDAR data, which in turn can be used for predicting aboveground carbon storage using a limited amount of field data at the plot level to inform policy and preservation policy and practice. The comparison/assessment of the models was part of a project titled “Assessing and Communicating Climate and Water Ecosystem Services of the City of Ann Arbor Greenbelt Program”. The student team did not have funding, so ARC consultants helped it get a free allocation from the Open Science Grid (OSG) to run the computations and a free storage allocation from the OSiRIS project for ‘cloud’ storage accessible from both UM and from OSG.