Jingpeng Wu joined the Flatiron Institute in 2020 as an associate research scientist to work on mapping neurons based on high-resolution Electron Microscopy images. Prior to coming here, Jingpeng was an associate research scholar at Princeton University, where he worked on petabyte-scale neuron reconstruction based on electron microscopy images using deep learning and cloud computing technologies. Jingpeng has a Ph.D. in biomedical engineering from Huazhong University of Science and Technology in China. In pursuing his Ph.D., he worked on large scale neuron and blood vessel tracing based on light microscopy images of whole mouse brains.
Principal Investigator: Dmitri Chklovskii
Fellow: Manuel Paez
U-Net is a U-shaped convolutional network architecture that is widely used in biological image segmentation. Convolutional networks are typical techniques in deep learning. The research group trains U-Net models to detect neuron membranes, synapses, mitochondria and glia, among other structures. Improving the model architecture with state-of-the-art techniques is expected to increase detection accuracy. The current U-Net is designed about four years ago and is outdated considering the rapid development of deep learning technologies. Recently, a combination of modern techniques has made the Convolutional Neural Networks (ConvNet) better than Transformers in terms of both speed and accuracy . Another direction of improvement is the loss function, we would like to improve the model with some state-of-the-art loss functions –. This project aims to test these features and incorporate the useful ones into the existing Deep Learning framework.