Honors Thesis
Computer science majors of junior or senior standing with at least a 3.3 grade point average in CS courses are eligible to join the Departmental Honors Program.
Requirements
In order to graduate with the Departmental Honors designation, a student must:
- Maintain at least a 3.3 grade point average in CS courses
- Have a Morrissey Contract Form approved by a faculty advisor and by the Honors committee by the end of their junior year (submit the forms to Mary Mulkeen in the CS main office)
- Complete two sections of CSCI4961 Honors Thesis during their senior year
- Submit a written honors thesis by the last day of class in the second semester of their senior year
- Make an oral presentation of their thesis at end of their senior year
Recent Honors Theses
Author: Steven Roche
Title: A 3-approximation Algorithm for the Min MaxCorrelation Clustering Problem
Abstract: In this paper, we give a 3-approximation algorithm to the min max correlation clustering problem. Given a complete graph, vertices are related by positive and negative edges. A positive edge denotes similar vertices and a negative edge denotes dissimilar vertices, and the goal is to minimize the l ∞-norm of disagreements over all vertices. The 3-approximation is possible by observing a structural property of vertices with degree greater than or equal to 3ø, where ø is our guess of the optimal solution. A combinatorial argument demonstrates the correctness of the algorithm, which identifies the optimal ω and has a runtime of O(n 2 D log D log n). This runtime includes determining the ω that corresponds to the optimal objective value over the graph.
Author: Michael Lin
Title: VesicleEM: Automated Segmentationand Classification Tools for Vesicles
Abstract: Vesicles are biological structures inside of neurons that facilitatecommunication between cells. Volume electron microscopy is widelyused for cellular imaging and enables researchers to detect vesicles’boundaries, content, and 3D structure. However, due to the high resolution of volumes and the comparatively large physical dimensions of neurons, many researchers face difficulties with segmentation, classification, and other spatial analysis of these electron microscopy datasets. We present VesicleEM (vEM) as a package for automated segmentation, classification, and proofreading of vesicles in electron microscopy images. Along with the machine learning based method for hands-off classification, we include a streamlined human proofreading tool. VesicleEM was designed and tested on a dataset of electron microscopy data of Hydra vulgaris; its performance is minute enough to generate unique vesicle instance segmentations along with a corresponding classification for each. In total, across a dataset of twenty neurons, we generate instance segmentations on 50043 unique vesicles. Each segmentation can be identified by a unique ID and 3D location, enabling further analysis. By combining image and morphology data for each vesicle, we can cluster vesicles into five unique types.
Author: Jason Ken Adhinarta
Title: Towards Foundation Models forBrain Microscopy Image Analysis
Abstract: Neuroscience imaging data presents unique challenges due to complex geometric structures andlimited data availability. Through three projects, we show how leveraging geometric priors,foundation models, and curated benchmarks are promising strategies for developing computervision techniques for neuroscience. First, using the Frenet-Serret Frame-based Decomposition,we study how invariances to curvilinear transformations induce data-efficient learning onpoint cloud part segmentation tasks. To verify our findings, we create CurviSeg, a syntheticdataset of 3D curvilinear structures, we curate DenSpineEM, comprised of 70 dendritessourced from public electron microscopy datasets, and we evaluate the generalizability ofour method on the IntrA intracranial aneurysm segmentation dataset. Secondly, we proposeTriSAM, adapting the Segment Anything Model for 3D segmentation of blood vessels usingtri-plane seed tracking. We demonstrate the effectiveness of our model on electron microscopydatasets sourced from mouse, macaque, and human cortical samples. Finally, we establishWormID-Bench, a benchmark for whole-brain neural activity extraction from C. elegans. Weassess how recent detection, identification, and tracking models perform, with the eventualgoal of promoting progress in reverse engineering the nervous system of the nematode. Wehope that this body of work advances understanding in neuroscience through the lens ofmachine learning methodology.