Direction 1: Real-Time Neural Signal Processing for Closed-Loop Neurofeedback Applications.
The miniaturized fluorescence microscope (Miniscope) and the tetrodes assembly are emerging techniques in observing the activity of a large population of neuros in vivo. It opens up new research opportunities for neuroscientific experiments in closed-loop for understanding how brain works. Recent years, the number of simultaneously recorded neurons has witnessed exponential increase. This project aims at creating high performance and energy efficient accelerators to support real-time neural signal processing for closed-loop neurofeedback applications.
We are working with leading team in this field at UCLA and focus on the backend real-time processing for the calcium imaging data collected from the head mounted miniature fluorescence microscope http://miniscope.org/. Technical challenges we are addressing include: Acceleration for real-time calcium image processing, which consists of pipelined stages including motion correction, image enhancement, neural activity extraction and decoding; Energy-efficient accelerator design and prototyping for miniaturized device with limited power supply.
Direction 2: Automation and Acceleration of Large-Scale Neural Reconstruction.
Recent advancements in microscopes allowing imaging an entire brain of a mouse in a couple days. However due to the intensive resolution and scale, computational processing of these images can take months. Although the results of these computational pipelines are analyzed offline after imaging, this processing can severely limits the workflow and hypothesis iteration speed. In our collaboration project with PIs Dr. William Yang (UCLA Neuroscience) and Dr. Hongwei Dong (USC Neuroscience) that was awarded an NIH U01 BRAIN grant, we accelerate such processing to study brain cell features of Huntington’s disease. The major run-time costs in these pipelines arise from converting dense three-dimensional images to a set of edges and nodes. These graphs offer a sparse and natural representation of neurons which have collections of branches whose pattern and connections determine their function or pathology. Graph processing algorithms are known for their acceleration difficulty due to their low computation to memory access ratio and their load-balance irregularity during parallelization. It is also increasingly more common for image pipelines to require additional specific image segmentation tasks such as cell body detection which are accomplished with more common neural network approaches. At a modest 30x magnification of a single mouse brain, the data size is still 20-30 teravoxels, therefore we focus on designs that optimize data-flow and prioritize computation at the edge, for example on the imaging workstation itself.
Broader impacts from these research projects shed light on establishing real-time closed-loop feedback control system, speeding up neural reconstruction of the entire brain of mouse, developing energy-efficient computing architecture, and accelerating the pace of neuroscience research and education worldwide.