[talks] Brian Matejek will present his MSE thesis "Learning Global Features for Neuron Reconstruction in EM Images" today, March 4, 2016 at 4:30pm in CS 401

[Nicki Gotsis]

Brian Matejek will present his MSE thesis "Learning Global Features for Neuron Reconstruction in EM Images" today, March 4, 2016 at 4:30pm in CS 401.

Tom Funkhouser (adviser) and Sebastian Seung are his committee members. 

All are welcome to join.  Please see thesis title and abstract below.

Title: Learning Global Features for Neuron Reconstruction in EM Images

Abstract:
Connectomics, the study of the neural connections, has gained significant
attention in the neuroscience community as a method to understand the
complex workings of the brain. Recent advancements in neural image
acquisition has created a surplus of data too expensive to manually
annotate leading to an influx of automatic reconstruction methods. Most
current methods employ hierarchical agglomeration techniques that focus on
extracted local features. These techniques fail to utilize overall global
structure of neurons and other valuable shape descriptors, leading to early
errors which compound over time. This paper introduces novel global and
local features for predicting merge candidates in oversegmented neural
images. Testing on the SNEMI3D challenge dataset, these new features
produce great improvement on traditional agglomeration algorithms.