/sym-STDP

Hao's sym-STDP-SNN model

Primary LanguagePython

Hao's sym-STDP-SNN model

update

Near-replication of Hao's SNN model from this paper using BindsNET.

Requirements

  • Ubuntu 16.04
  • Python 3.6
  • BindsNET 0.2.4

N.B. Versions mentioned above are verified to work together. Other versions might not be compatible.

Setup

Head over to BindsNET and follow their instruction for environment setup.

Alternatively, install the required packages using requirements.txt provided:

pip install -r requirements.txt

Run

Simply run the following command (both training and inference process included in the file):

python main_snn.py

Example of passing optional command-line arguments:

# (--n_neurons [int]): to set number of excitatory, inhibitory neurons
# (--gpu): to use GPU backend for calculation
python main_snn.py --n_neurons 400 --gpu

Run the script with the --help or -h flag for more information.

Models

  • v1

    Original model used in the paper but was not able to replicate the results.

  • v2 (Current default model used in main_snn.py.)

    Modified from original (v1) model. Removes inhibitory layer and replaces it with a recurrent inhibitory connection in the excitatory layer.

  • v3

    Based on v2 model. Modified the weight normalization between hidden layer -- SL layer to normalize out-synapses of neurons in SL layer. (Originally normalize in-synapses in the paper.)

Results

During inference stage, there's a problem of SL neurons not spiking due to insufficient response towards the input sample, and thus can't infer any label for the input sample. To counter this problem, added a function (named re-input in below table) to re-present samples with increased firing rates (Hz) till SL neurons fires at least a spike. (Similar function for excitatory neurons during training stage.)

Note: This function was not mentioned in the paper.

Below table shows approx. test accuracies for each model with or without the function (named re-input) mentioned above used. Results of 100 excitatory neurons model trained with 1 epoch:

Model w/o re-input (%) w/ re-input (%)
v2 72.53 80.14
v3 81.52 82.09

Citation

  • BindsNET

    All of the SNN models in this project were implemented using BindsNET:

    Link to GitHub.

    @ARTICLE{10.3389/fninf.2018.00089,
	AUTHOR={Hazan, Hananel and Saunders, Daniel J. and Khan, Hassaan and Patel, Devdhar and Sanghavi, Darpan T. and Siegelmann, Hava T. and Kozma, Robert},
	TITLE={BindsNET: A Machine Learning-Oriented Spiking Neural Networks Library in Python},
	JOURNAL={Frontiers in Neuroinformatics},
	VOLUME={12},
	PAGES={89},
	YEAR={2018},
	URL={https://www.frontiersin.org/article/10.3389/fninf.2018.00089},
	DOI={10.3389/fninf.2018.00089},
	ISSN={1662-5196},
}

  • sym-STDP-SNN

    Some of the parameters' value that are not mentioned in the paper were obtained from Hao's original code.

    Link to GitHub.

     @ARTICLE{hao2019biologically,
 	AUTHOR={Hao, Yunzhe and Huang, Xuhui and Dong, Meng and Xu, Bo},
 	TITLE={A biologically plausible supervised learning method for spiking neural networks using the symmetric STDP rule},
 	JOURNAL={Neural Networks},
 	VOLUME={121},
 	PAGES={387-395},
 	YEAR={2020},
 	PUBLISHER={Elsevier},
 	URL={https://www.sciencedirect.com/science/article/pii/S0893608019302680},
 	DOI={10.1016/j.neunet.2019.09.007},
 }