oscarliau / sparsenet

Uniform Sparse Network (USN) and Sparse Neural Networks

Citation

Please cite the USN paper when using this open-source package. This will help the project and our team:

Luo W. Improving Neural Network with Uniform Sparse Connectivity. IEEE Access, 2020, accepted, doi: 10.1109/ACCESS.2020.3040943

@article{luo:2020,
  author =    {Weijun Luo},
  title =     {Improving Neural Network with Uniform Sparse Connectivity},
  journal =   {IEEE Access},
  year =      2020,
  volume =    {8},
  number =    {},
  pages =     {}
}

Overview

The novel uniform sparse network (USN) has even and sparse connectivity within each layer (or uniform sparse layer). USN models has one striking property: at given connection density, USN topology can differ a lot (Figure 1 below) but performance stays the same. Unlike other sparse neural network models, no structure search or optimization is needed for USN models. USN:

• is a natural generalization of the fully connected network (FCN);
• is static sparse neural network, search-free, robust and fast to train;
• can improve FCN in accuracy, robustness and scalability simultaneously by tunning the connection density/sparsity hyperparameter;
• is robust against overfitting, can easily scale up to big neural network without extra regularization (eg. dropout, L1 or L2 etc);
• shows the potential to fully replace FCN in a wide range of deep learning architectures and broad applications.

USN and random sparse networks are implemented in sparsenet as options of sparse layer class (USN is the default). sparse layer is very similar to the commonly used Dense layer in Keras and tf.keras in both coding and function, and intends to be a full replacement of Dense layer in all aspects and conditions. Just like Dense layer, sparse layer can be used as basic building block for complex neural networks and deep learning models.

USN definition (Figure 1)

USN with CNN and RNN architecture for MNIST dataset (Figure 2)

Installation (unix/linux/bash command line)

you can pip install sparsenet package from GitHub. This will also install the dependencies, eg TensorFlow and NumPy.

pip install git+https://github.com/datapplab/sparsenet

Alternatively, you can download sparsenet package from GitHub. Note you need to have Python 3 (and NumPy) and TensorFlow 2.0 or above installed first.

# pip install --upgrade tensorflow
cd /path/to/your/local/directory/
git clone https://github.com/datapplab/sparsenet.git

Quick start (Python code)

As shown in the quick example below, sparse layer is used the same way as Dense layer in Keras, except with a few additional arguments like density (connection density or sparsity). Please check the documentation with the code for details.

#import tensorflow
import tensorflow as tf

#import sparse
#need to specify the local path if sparsenet was downloaded/cloned instead of installed
#import sys
#sys.path.append('/path/to/your/local/directory/sparsenet')
from sparsenet.core import sparse

#define a simple two-layer neural network (MLP) with USN layers
#alternatively, you can use Dense layers as in the lines commented out
model = tf.keras.models.Sequential([
  # X_train is your training data (dim 0: samples and dim 1: features)
  tf.keras.layers.Input(shape=X_train.shape[1]),
  # tf.keras.layers.Dense(128, activation=None),
  sparse(units=128, density=0.1, activation=None),
  # tf.keras.layers.Dense(10, activation='softmax')
  sparse(units=10, density=0.5, activation='softmax')
])

#sparse layer is USN by default, to specify random sparse network/layer instead:
  sparse(units=128, density=0.1, activation=None, confun="random_2d")

Tutorials with Full examples

In this repository, you can find two tutorials on USN and sparse neural networks.

• Tutorial on basic sparse neural networks with scRNA-Seq data
• Tutorial on basic and deep sparse neural networks with MNIST data

More information

Please check the tutorials, package documentation and the USN paper for more info.

You can reach the author at luo_weijun[AT]yahoo.com

Thank you for your interest.