αCSC

This is a library to perform shift-invariant dictionary learning on neural time-series data. Framed mathematically, if we are giving a signal \(x_n \in \mathbb{R}^T\), we want to learn \(k\) atoms \(d^k \in \mathbb{R}^{L}\) and their associated activations \(z^k_n \in \mathbb{R}^{T - L + 1}\). The optimization problem boils down to minimizing an \(\ell_2\) reconstruction loss with an \(\ell_1\) penalty term, i.e.,

\[\min_{d,z} \sum_n (\|x_n - \sum_k d^k * z^k_n \|_2^2 + \lambda \sum_k z_n^k)\]

subject to \(z_k >= 0\). The shift invariance is encoded by the convolution operator \(*\) which is why these methods are called “convolutional sparse coding”.

The alpha in alphaCSC [1] refers to the fact that we assume an alpha-stable distribution for the noise. This leads to a weighted formulation of `vanilla’ CSC, the weights being used to downweight noisy portions of the data. Or in other words:

\[\min_{d,z} \sum_n (\|\sqrt{w_n} \odot (x_n - \sum_k d^k * z^k_n) \|_2^2 + \lambda \sum_k z_k)\]

where \(w_n\) are the weights which are learned using an EM algorithm.

Installation

We recommend the Anaconda Python distribution. To install alphacsc, you first need to install its dependencies:

$ pip install numpy matplotlib scipy

Then install alphacsc:

$ pip install git+https://github.com/alphacsc/alphacsc.git#egg=alphacsc

If you do not have admin privileges on the computer, use the --user flag with pip. To upgrade, use the --upgrade flag provided by pip.

To check if everything worked fine, you can do:

$ python -c 'import alphacsc'

and it should not give any error messages.

Quickstart

All you need is a numpy array X of shape n_trials x n_times. You need to specify the length of atoms n_times_atom and the number of atoms n_atoms you would like to estimate. The penalty parameter reg controls the sparsity in the activations estimated.

>>> from alphacsc import learn_d_z
>>> n_atoms, n_times_atom, n_iter = 2, 64, 60
>>> pobj, times, d_hat, Z_hat = learn_d_z(X, n_atoms, n_times_atom, reg=reg, n_iter=n_iter)  

Bug reports

Use the github issue tracker to report bugs.

Cite

[1] Jas, M., Dupré La Tour, T., Şimşekli, U., & Gramfort, A. (2017). Learning the Morphology of Brain Signals Using Alpha-Stable Convolutional Sparse Coding. arXiv preprint arXiv:1705.08006.