EnsembleSTDP

This is the source code for the paper:

• Title: EnsembleSTDP: Distributed in-situ Spike Timing Dependent Plasticity Learning in Spiking Neural Networks. [PDF]
• Authors: Hanyu Yuga and Khanh N. Dang
• Venue: 2024 IEEE 17th International Symposium on Embedded Multicore/Many-core Systems-on-Chip (MCSoC), Dec. 16-19, 2024.
• URL: https://ieeexplore.ieee.org/abstract/document/10819516

Requirements

Python <3.11,>=3.8.1

Set up

Below are the steps to set up the environment for this package.

Clone repository

git clone https://github.com/klab-aizu/EnsembleSTDP

Install bindsnet

1. Change the directory to the top level of EnsembleSTDP
2. Clone the BindsNet repository

git clone https://github.com/BindsNET/bindsnet

3. Change the directory to the top level of BindsNet
4. Make bindsnet editable

pip install -e .

5. Go back to the top level of EnsembleSTDP
6. Move the modified models.py to the BindsNet library

mv models.py bindsnet/bindsnet/models/models.py

How to use scripts

Here is how to perform basic operations in ensemble learning

Train sub-model

An SNN sub-model is defined by the following attributes:

• model_number
• n_neurons (number of excitatory neurons)
• starting index of training images
• ending index of training images

For example, to train a sub-model 1 that has 100 neurons and trained from 0 to 6000 images, run:

python eth_mnist_train.py --model_number 1 --n_neurons 100 --start_image 0 --end_image 6000

Test sub-model

After the training, the sub-model’s classification accuracy can be tested by replacing "train" with "test" in the above command

python eth_mnist_test.py --model_number 1 --n_neurons 100 --start_image 0 --end_image 6000

Merge sub-models

To merge and compress the trained sub-models, the number of compressions and similarity measurements need to be specified. Before running the command below, make sure all the sub-model files to be merged (model_.pth, assignments_.pth, and proportions_*.pth for each model) are in the models_to_merge folder. The example of the command to merge sub-models using MSE and compress 100 neurons:

python3 eth_mnist_merge.py --method mse --n_compression 100

• Mean Squared Error (MSE) -> --method mse
• Manhattan Distance -> --method manhattan
• Cosine Similarity -> --method cos
• Correlation Coefficient -> --method corr

Plot 2D-Receptive field

To visualize the trained weights of a sub-model, run the following command. The example of visualizing sub-model 1 with 100 neurons, trained from 0 to 6000 images

python eth_mnist_plot.py --model_number 1 --n_neurons 100 --start_image 0 --end_image 6000

The receptive field will be saved as receptive_field_model_1_100_neurons_0_to_6000.png

*note that models trained on GPU cannot be plotted on CPU, or vice versa

Training or testing multiple sub-models at once

There is a supplemental shell script to run the training, testing, and merging on a series of command line arguments. For example, to train the following 5 sub-models at once,

• model_1_100_neurons_0_to_6000 (model 1 with 100 neurons trained from 0 to 6000 images)
• model_2_100_neurons_6000_to_12000
• model_3_100_neurons_12000_to_18000
• model_4_100_neurons_18000_to_24000
• model_5_100_neurons_24000_to_30000

first, set the parameters as follows in input.txt
1 0 6000
2 6000 12000
3 12000 18000
4 18000 24000
5 24000 30000

Make sure to add a new line after the final arguments

then edit shell.sh

#!/bin/bash
input_file="input_args.txt"
output_file="output_results.txt"

# Run program for each set of arguments in the input file
while read -r model_number start_image end_image ; do
    python eth_mnist_train.py --model_number "$model_number" --start_image "$start_image" --end_image "$end_iamge" --n_neurons 100 >> "$output_file"
done < "$input_file"

Finally, run

./shell.sh

Contact

Khanh N. Dang: khanh@u-aizu.ac.jp