Supervised Machine Learning for Greedy Agglomeration in Connectomics

Abstract:

This study explores the use of supervised machine learning methods for greedy agglomeration
in the application of constructing connectomes or neural wiring diagrams
that show how neurons are connected to each other. The current approach
to this problem is mean affinity greedy agglomeration, which makes locally optimal
merge/not-merge decisions. The rationale behind using supervised learning methods
is that they may lead to locally optimal merge/not-merge decisions that are more
globally optimal than those made by mean affinity agglomeration. The results of this
study are inconclusive on whether supervised machine learning methods are better
than mean affinity agglomeration because of evaluation issues. Future work should
thus explore better methods of evaluating agglomeration performance. In addition,
future work should address the specific weaknesses of the pipeline for producing connectomes:
misalignment errors and errors where the dendritic spines do not grow out
from the dendritic shafts.