NALSM: Neuron-Astrocyte Liquid State Machine
This package is a Tensorflow implementation of the Neuron-Astrocyte Liquid State Machine (NALSM) that introduces astrocyte-modulated STDP to the Liquid State Machine learning framework for improved accuracy performance and minimal tuning.
The paper has been accepted at NeurIPS 2021, available here.
Citation
Vladimir A. Ivanov and Konstantinos P. Michmizos. "Increasing Liquid State Machine Performance with Edge-of-Chaos Dynamics Organized by Astrocyte-modulated Plasticity." 35th Conference on Neural Information Processing Systems (NeurIPS 2021).
@inproceedings{ivanov_2021,
author = {Ivanov, Vladimir A. and Michmizos, Konstantinos P.},
title = {Increasing Liquid State Machine Performance with Edge-of-Chaos Dynamics Organized by Astrocyte-modulated Plasticity},
year = {2021},
pages={1--10},
booktitle = {35th Conference on Neural Information Processing Systems (NeurIPS 2021)}
}
Software Installation
- Python 3.6.9
- Tensorflow 2.1 (with CUDA 11.2 using tensorflow.compat.v1)
- Numpy
- Multiprocessing
Usage
This code performs the following functions:
- Generate the 3D network
- Train NALSM
- Evaluate trained model accuracy
- Evaluate trained model branching factor
- Evaluate model kernel quality
Instructions for obtaining/setting up datasets can be accessed here.
Overview of all files can be accessed here.
1. Generate 3D Network
To generate the 3D network, enter the following command:
python generate_spatial_network.py
This will prompt for following inputs:
WHICH_DATASET_TO_GENERATE_NETWORK_FOR? [TYPE M FOR MNIST/ N FOR NMNIST]: enterMto make a network with an input layer sized for MNIST/Fashion-MNIST orNfor N-MNIST.NETWORK_NUMBER_TO_CREATE? [int]: enter an integer to label the network.SIZE_OF_LIQUID_DIMENSION_1? [int]: enter an integer representing the number of neurons to be in dimension 1 of liquid.SIZE_OF_LIQUID_DIMENSION_2? [int]: enter an integer representing the number of neurons to be in dimension 2 of liquid.SIZE_OF_LIQUID_DIMENSION_3? [int]: enter an integer representing the number of neurons to be in dimension 3 of liquid.
The run file will generate the network and associated log file containing data about the liquid (i.e. connection densities) in sub-directory
/
/networks/
.
2. Train NALSM
2.1 MNIST
To train NALSM model on MNIST, enter the following command:
python NALSM_RUN_MAIN_SIM_MNIST.py
This will prompt for the following inputs:
GPU?: enter an integer specifying the gpu to use for training.VERSION? [int]: enter an integer to label the training simulation.NET_NUM_VAR? [int]: enter the number of the network created in Section 1.BATCH_SIZE? [int]: specify the number of samples to train at same time (batch), for liquids with 1000 neurons, batch size of 250 will work on a 12gb gpu. For larger liquids(8000), smaller batch sizes of 50 should work.BATCHS_PER_BLOCK? [int]: specify number of batchs to keep in memory for training output layer, we found 2500 samples works well in terms of speed and memory (so for batch size of 250, this should be set to 10 (10 x 250 = 2500), for batch size 50 set this to 50 (50 x 50 = 2500).ASTRO_W_SCALING? [float]: specify the astrocyte weight detailed in equation 7 of paper. We used 0.015 for all 1000 neuron liquids, and 0.0075 for 8000 neuron liquids. Generally accuracy peaks with a value around 0.01 (See Appendix).
This will generate all output in sub-directory
/
/train_data/ver_XX/
where
XX is
VERSION number.
2.2 N-MNIST
To train NALSM model on N-MNIST, enter the following command:
python NALSM_RUN_MAIN_SIM_N_MNIST.py
All input prompts and output are the same as described above for run file NALSM_RUN_MAIN_SIM_MNIST.py.
2.3 Fashion-MNIST
To train NALSM model on Fashion-MNIST, enter the following command:
python NALSM_RUN_MAIN_SIM_F_MNIST.py
All input prompts and output are the same as described above for run file NALSM_RUN_MAIN_SIM_MNIST.py.
Instructions for training other benchmarked LSM models can be accessed here.
3. Evaluate Trained Model Accuracy
To get accuracy of a trained model, enter the following command:
python get_test_accuracy.py
The run file will prompt for following inputs:
VERSION? [int]: enter the version number of the trained model
This will find the epoch with maximum validation accuracy and return the test accuracy for that epoch.
4. Evaluate Model Branching Factor
To compute the branching factor of a trained model, enter the following command:
python compute_branching_factor.py
The run file will prompt for following inputs:
VERSION? [int]: enter the version number of the trained model.
The trained model directory must have atleast one .spikes file, which contains millisecond spike data of each neuron for 20 arbitrarily selected input samples in a batch. The run file will generate a .bf file with same name as the .spikes file.
To read the generated .bf file, enter the following command:
python get_branching_factor.py
The run file will prompt for following inputs:
VERSION? [int]: enter the version number of the trained model.
The run file will print the average branching factor over the 20 samples.
5. Evaluate Model Kernel Quality
Model liquid kernel quality was calculated from the linear speration (SP) and generalization (AP) metrics for MNIST and N-MNIST datasets. To compute SP and AP metrics, first noisy spike counts must be generated for the AP metric, as follows.
To generate noisy spike counts for NALSM model on MNIST, enter the following command:
python NALSM_RUN_MAIN_SIM_MNIST_NOISE.py
The run file requires a W_INI.wdata file (the initialized weights), which should have been generated during model training.
The run file will prompt for the following inputs:
GPU?: enter an integer to select the gpu for the training simulation.VERSION? [int]: enter the version number of the trained model.NET_NUM_VAR? [int]: enter the network number of the trained model.BATCH_SIZE? [int]: use the same value used for training the model.BATCHS_PER_BLOCK? [int]: use the same value used for training the model.
The run file will generate all output in sub-directory
/
/train_data/ver_XX/
where
XX is VERSION number.
To generate noisy spike counts for NALSM model on N-MNIST, enter the following command:
python NALSM_RUN_MAIN_SIM_N_MNIST_NOISE.py
As above, the run file requires 'W_INI.wdata' file. All input prompts and output are the same as described above for run file NALSM_RUN_MAIN_SIM_MNIST_NOISE.py.
After generating the noisy spike counts, to compute the SP and AP metrics for each trained model enter the following command:
python compute_SP_AP_kernel_quality_measures.py
The run file will prompt for inputs:
VERSION? [int]: enter the version number of the trained model.DATASET_MODEL_WAS_TRAINED_ON? [TYPE M FOR MNIST/ N FOR NMNIST]: enter dataset the model was trained on. The run file will print out the SP and AP metrics.
Instructions for evaluating kernel quality for other benchmarked LSM models can be accessed here.
2.5k Dec 30, 2022
2 Dec 13, 2021
743 Jan 01, 2023
129 Dec 13, 2022
152 Oct 27, 2022
332 Jan 03, 2023
7 Jul 20, 2021
1.9k Jan 09, 2023
30 Nov 19, 2022
10 Oct 19, 2022
1.4k Jan 07, 2023
1.2k Jan 03, 2023
109 Dec 29, 2022
55 Dec 06, 2022
258 Jan 07, 2023
423 Jan 02, 2023
14 Nov 20, 2022
687 Jan 04, 2023
0 Jun 08, 2022
1 Nov 12, 2021