Localized Sparse Coding

Background

Learning sparse codes from compressed representations with biologically plausible local wiring constraints

Code accompanying 2020 NeurIps paper by Kion Fallah^*, Adam A. Willats^*, Ninghao Liu, and Christopher J. Rozell.

This library provides code to learn a sparse coding model with different coefficient inference strategies (FISTA and ADMM), as well as different compression schemes (dense random matrix, block diagonal matrix, and banded random matrix).

_{* equal contribution}

Usage

Training

In order to train a compressed sparse dictionary, use the train_sparse_dict.py script. There are several optional parameters, but the ones that you might be interested in are:

• --compression or -c, which determines the compression matrix. Options are none for no compression, bdm for banded diagonal matrix, and brm for banded random matrix. For a dense random, use bdm with a localization of 1, -l 1.
• --localization or -l, which determines the degree of localization

Example usage:

python train_sparse_dict.py -c bdm -l 4

Dependencies

Python 3.0+, Scikit-Learn, Numpy, Scipy.

Whitened natural images retrieved from http://www.rctn.org/bruno/sparsenet/.

Examples

Output

The training script, train_sparse_dict.py, outputs a dictionary containing Numpy arrays in a specific format. To use it, first load a training file:

data_file = np.load('./results/traindata_05-31-2020_none_J1.npz')

With any training file loaded, access the dictionary using the following format:

Dictionary key	Value
data_file['phi']	A Numpy tensor containing the dictionary at each epoch in training. Dimensions [num_epochs x patch_size**2 x dict_count]
data_file['time']	The training time, in seconds, at each epoch Dimensions [num_epochs x float]
data_file['train_loss']	Training loss at each epoch. Performed on uncompressed or compressed dictionaries learned on the main data-set. Loss is MSE reconstructing training patches. Dimensions [num_epochs x float].
data_file['val_loss']	Validation loss at each epoch. Performed on uncompressed or reconstructed dictionaries on hold-out data-set, taken from seperate images. Loss is MSE between reconstructed validation patches. Dimensions [num_epochs x float].

Uncompressed

Distinct Block Diagonal Matrix, L = 1/4

Banded Random Matrix, L = 1/4