High resolution neural connectivity from incomplete tracing data using nonnegative spline regression
Kameron Decker Harris (kamdh@uw.edu), Stefan Mihalas (stefanm@alleninstitute.org), Eric Shea-Brown (etsb@uw.edu).
NIPS, 2016
The paper is available at https://nips.cc, here, and at https://arxiv.org/abs/1605.08031
The majority of code is split into separate repositories:
- allen-voxel-network - utilities for setting up voxel matrices by pulling from allensdk
- spatial-network-regression - solves (P1) using L-BFGS-B
Furthermore, we provide here the MATLAB code used to solve (P2), the low-rank version, using projected gradient descent:
- proj_grad_low_rank.m
Projections from a source voxel in VISp, depicted in blue, to the rest of the visual areas. The main discrepancy between the full and low rank solutions is confined to the medial-posterior area of VISp. There, the low rank solution undershoots the full rank solution in proximal projections, and overshoots it with distal projections.
- region_names.png - 2-D projection of region labels
- movie_full.mp4 - solution of (P1), lambda=10^5
- movie_low_rank.mp4 - solution of (P2), lambda=10^5, r=160
- movie_res.mp4 - residual (W_lowrank - W_full) plotted in the same way
- movie_full_retrograde.mp4 - solution of (P1), lambda=10^5, visualized retrograde with W^T