FastSRM

Implementation of FastSRM algorithms.

The model of probabilistic SRM is given by:

where

• is the data of subject
• is the basis of subject
• is the shared response (or sources) assumed to be sampled from a centered Gaussian with covariance
• is the noise in subject assumed to be sampled from a centered Gaussian with covariance where is the identity matrix. We call the noise variance of subject .

In practice we observe n samples of . When the number of samples is much lower than the number of features v, the SRM model can be fitted efficiently. This is what this repository provides. We also assume that the covariance of the shared response is diagonal to obtain identifiability.

Install

Clone the repository

git clone https://github.com/hugorichard/FastSRM

In the cloned directory execute

pip install -e .

Requirements:

scipy>=0.18.0
numpy>=1.12
scikit-learn>=0.23
joblib>=1.1.0
matplotlib>=2.0.0
pytest>=6.2.5

To run the experiments on identifiability, brainiak (https://github.com/brainiak/brainiak) and nilearn (https://nilearn.github.io/stable/index.html) are necessary.

Usage

In many neuroscience datasets, the samples are split into sessions. Therefore, for each subject we will have several sessions that can be time-wise concatenated to obtain all samples. In IdentifiableFastSRM

# Input data X: neuroimaging data 
# X is a np array of shape (n_subjects, n_sessions)
# X[i, j] is a path to a np array of shape (n_voxels, n_timeframes)
from fastsrm.identifiable_srm import IdentifiableFastSRM
srm = IdentifiableFastSRM(n_components=5, temp_dir="./", n_jobs=5)
S = srm.fit_transform(X) # Shared response: np array of shape (n_components, n_timeframes)
W = srm.basis_list # Shared response: np array of shape (n_components, n_timeframes)
Sigma = srm.source_covariance # (Diagonal) Covariance of the shared response: np array of shape (n_components,)
sigmas = srm.noise_variance # Variance of the noise: np array of shape (n_subjects)

Run experiments

Synthetic experiment

1. In the experiments folder run ipython synthetic.py

2. Go in the plotting folder and run ipython plot_synthetic_gradients.py

The figure is now available in the figures folder under the name synthetic_gradient.pdf

Timesegment matching experiment

1. Go in the experiments folder and download the sherlock dataset by running: ipython get_sherlock.py

2. Go in the experiments folder and run bash timesegment.bash if you have access to a cluster with slurm otherwise run bash timesegment_local.bash

3. Go in the plotting folder and run ipython plot_timesegment.py

The figure is now available in the figures folder under the name timesegment_matching_sherlock.pdf

Identifiability experiment

1. Go in the experiments folder and download the sherlock dataset by running: ipython get_sherlock.py

2. Download and install Brainiak (see https://github.com/brainiak/brainiak)

3. Go in the experiments folder and run bash identifiability.bash if you have access to a cluster with slurm otherwise run bash identifiability_local.bash

4. Go in the plotting folder and run ipython plot_identifiability.py

The figure is now available in the figures folder under the name identifiability.pdf

Documentation

https://hugorichard.github.io/FastSRM/