Purpose

This repo gathers the required data and code to reproduce the neural stitching results in the published LFADS paper, specifically those results presented in Figure 4.

Pandarinath C, O’Shea DJ, Collins J, Jozefowicz R, Stavisky SD, Kao JC, et al. Inferring single-trial neural population dynamics using sequential auto-encoders. Nat Methods. 2018;15: 805–815. doi:10.1038/s41592-018-0109-9 https://www.nature.com/articles/s41592-018-0109-9

This repo was prepared by Daniel O'Shea, for questions or issues, it would be preferable to file an issue on Github directly.

What you'll need

• Git LFS (required to download the larger portions of this repo): https://git-lfs.github.com
• MATLAB, particularly a relatively modern version. I'm using 2019a but 2017 onwards should work.
• LFADS Run Manager: https://github.com/lfads/lfads-run-manager (see docs at https://lfads.github.io/lfads-run-manager/), if you want to prepare an LFADS run and load the results into MATLAB easily
• LFADS Python / TensorFlow, if you want to retrain the LFADS model rather than use the exported posterior means: see https://lfads.github.io/lfads-run-manager/install/ for installation instructions

Clone the repo

Once you've downloaded Git LFS and installed it, you should be able to run:

git lfs clone https://github.com/nplcode/lfads-neural-stitching-reproduce.git

Details

If you just want the data, the data fed into LFADS are located in export_v05_broadbandRethreshNonSorted_filtered as mat files and the outputs (LFADS posterior means) have been exported as HDF5 files into posterior_means_export for all of the single dataset runs and the stitched model.

If you'd like to rerun LFADS yourself, follow along in the notebook provided neural_stitching_walkthrough.ipynb. Then you can train LFADS with instructions provided at https://lfads.github.io/lfads-run-manager/.

The decoding and analysis scripts are too interwoven into my personal analysis toolkit to extract easily, but I've included the code regardless for your reference. The saved outputs from these scripts are included in the results folder. Some of the scripts inside +PierreEricLFADS reference these saved outputs to generate individual figure panels directly, and thus do not depend on the kinematic decoding. The primary result decoding kinematics from LFADS factors is done in kinematicDecodeFromFactors.m.

Questions?

Please file an issue on Github first unless you need to contact me directly at djoshea at stanford. Thanks!