eden-kramer-lab/replay_identification

A semi-latent state-space model that combines movement, LFP, and ensemble spiking information to identify periods of replay and decode its content in real time. Long Tao et al., unpublished.

replay_identification

A semi-latent state-space model that combines movement, LFP, and ensemble single unit/multiunit information to identify periods of replay and decode its content.

NOTE: This code is still in production and prepublication.

Installation

replay_identification can be installed through pypi or conda. Conda is the best way to ensure that everything is installed properly.

pip install replay_identification
python setup.py install

Or

conda install -c edeno replay_identification
python setup.py install

Usage

See the notebooks (#1, #2) for more information on how to use the package.

You can also use the launch binder button at the top of the Readme to play with simulated data in your web browser.

Package Requirements

• numpy
• scipy
• statsmodels
• numba
• matplotlib
• xarray
• scikit-learn
• regularized_glm

See the setup.pyor environment.yml file for the most up to date list of dependencies.

Developer Installation

1. Install miniconda (or anaconda) if it isn't already installed. Type into bash (or install from the anaconda website):

wget https://repo.continuum.io/miniconda/Miniconda3-latest-Linux-x86_64.sh -O miniconda.sh;
bash miniconda.sh -b -p $HOME/miniconda
export PATH="$HOME/miniconda/bin:$PATH"
hash -r

2. Go to the local repository (.../replay_identification) and install the anaconda environment for the repository. Type into bash:

conda update -q conda
conda info -a
conda env create -f environment.yml
source activate replay_identification
python setup.py develop

replay_identification's Issues

Refit place fields to local states

Speed up prediction of LFP power

The slowest part of prediction is computing the lfp power likelihood. Maybe shift this to parametric or figure out a faster way than using statsmodels KDEMultivariate

Add estimation of initial conditions

Changing the bin size causes inaccurate estimation by the smoother

There is probably a missing multiplication by bin size somewhere

Implement Viterbi for finding non-local states

Make API more like scikit-learn

Speed up fitting of spiking models

Slowest part of the fitting step is fitting the place field models. Part of this is due to using L2 regularization in statsmodels.

Implement filter and smoother using `log_sum_exp` trick to prevent under and overflow

Should the no replay likelihood be evaluated by bin instead of exact position

Would avoid having to recompute the likelihood twice.

Include fragmented state

Evaluate if speed model is actually informative

Double check fitting of state transitions