A Spotlight on What to Learn for Neural Reconstruction from EM Data
Abstract: The goal of electron microscopic (EM) neural reconstruction, or connectomics, is to discover the neural wiring diagram from images of animal brain tissue. The computational framework for EM reconstruction typically consists of machine vision tasks such as volume segmentation and synaptic connectivity detection.
In this talk, I will present insights gathered from multiple connectomic reconstructions from EM images collected from different animals with different imaging techniques. In particular, I will emphasize defining target functions (analytical or nonanalytical mapping) for learning segmentation and synapse detection to achieve optimal performance. The presentation will include quantitative and qualitative results from methods designed for different connectomic tasks as well as some discussion of large-scale computation and potential future directions.