Sergiy Popovych will present his Pre FPO " Visual Feature Extraction with Unsupervised Metric Learning" on Friday August 7th, 2020 at 8:30pm ET via Zoom. Zoom Link: [ https://princeton.zoom.us/j/94304733448 | https://princeton.zoom.us/j/94304733448 ] The members of his committee are as follows: Sebastian Seung (Advisor); Readers: Sebastian Seung and Viren Jain (Google); Examiners: Stephan Saalfeld (Janelia), Yoram Singer, and Kai Li Please see abstract below. Abstract: In serial section electron microscopy (ssEM), a brain volume is cut into a series of ultrathin sections, each of which is imaged via electron microscopy. Then the 2D images are combined to create a 3D image stack, from which the connections between neurons can be reconstructed. Images obtained from serial section electron microscopy often contain distortions of both affine and non-affine nature. Before the 3D stack can be analyzed, these tissue distortions have to be corrected. The problem of correcting these distortions, also known as 3D image registration or stack alignment, poses significant challenges for ssEM neuron reconstruction. This thesis focuses on convolutional neural network (ConvNet) based method that can successfully align large stacks, even in presence of large discontinuous deformations. The key behind the proposed method is a metric learning based technique that learns to extract dense noiseless representation of ssEM images. In this technique, two ConvNets are trained simultaneously. The first ConvNet is a feature extractor that is applied to the images independently. The extracted features are passed to the second ConvNet, which predicts a corrective deformation field. The joint loss of the two networks is to minimize the mean square error between the extracted features after the corrective field is applied, while maximizing the difference between the features before the corrective field is applied. Special techniques are employed to encourage feature density and to avoid degenerate solutions. The proposed technique is evaluated on real world datasets and is shown to outperform other existing approaches. This thesis also presents an efficient distributed pipeline that can be used to efficiently apply the proposed method to multi-terabyte datasets.