RAVSim v2.0: Run-Time Analysis and Visualization Simulator for SNN

under-review

* Corresponding Author:

Paper (under-review)

Code

Simulator

Supplementary

Abstract

RAVSim, short for Run-time Analyzing and Visualization Simulator, is an interactive tool designed for the analysis and simulation of spiking neural network models. Developed on the LabVIEW (Laboratory Virtual Instrument Engineering Workbench) platform, RAVSim uses a multi-core architecture to provide users with a versatile environment for exploring the behavior of spiking neural networks.



What is RAVSim Offering

RAVSim offers both deterministic (by solving Ordinary Differential Equations (ODEs) and stochastic (using stochastic leaky integrate-and-fire (LIF) neurons algorithm) simulations. This dual simulation capability enables users to gain a comprehensive understanding of spiking neural network dynamics through various approaches. What sets RAVSim apart from other tools is its unique ability to execute, analyze, extract, and validate models using image-based datasets. This empowers users not only to analyze and simulate spiking neural networks but also to create custom datasets tailored to their specific image pixel, quality, and extension preferences.

A Run-time VI

A run-time VI offers three different WTA Networks, List of conditions:
1- Fully connected network
2- source != target index neurons
3- source == target index neurons
Also, a mixed signal plot for a better understanding of neuron communication.




Dataset Preprocessing VI

Within dataset preprocess VI, users are allowed to use downloaded images for creating a scalable dataset. This flexibility includes the ability to use images with varying extensions, sizes, and qualities, pixel to the user's specific requirements. The tool supports a wide variety of image formats, accommodating datasets that suit diverse research and application needs.





Image Classification VI

RAVSim has recently introduced new features to enhance the user experience. It now offers a convenient option to generate weights for your spiking neural network models using image classification training. This feature eliminates the need for complex coding; you can simply utilize RAVSim to train your model and generate the required weights effortlessly.





SNN Models Comparison VI

In addition to weight generation, RAVSim facilitates the comparison of different spiking neural network models. Users can dynamically update sets of parameter values for each model at runtime and obtain comparative results. This functionality aids in identifying the most suitable model for your specific spiking neural network application.





RAVSim Introduction






Practical Demonstration

The RAVSim is an open-source simulator and it is available publicly at "Here" (RAVSim v2.0 is under-review by LABView team). To learn how to effectively use this tool at runtime experiments, this document document provides a comprehensive guide on using RAVSim for performing run-time experiments.:

RAVSim v1.0 Demonstration


YouTube Source





RAVSim v2.0 Demonstration


YouTube Source




Qualitative Results

Neural Model interactions during runtime

The runtime interactive environment is designed to facilitate real-time adjustments to parameter values during the simulation. This live modification capability empowers users to observe the immediate effects of parameter changes, enhancing their understanding of how different variables impact the behavior of the spiking neural network.
Moreover, when instantaneous triggering of input model concentrations to the threshold level is necessary, the neuron communication plot visually presents the real-time excitation results of the neurons.
A run-time VI offers three different WTA Networks, List of conditions:
* Fully connected network.
* source != target index neurons.
* source == target index neurons.
* Also, a mixed signal plot for a better understanding of neuron communication.

input


* On left side - A runtime interactive environment.
* Center - A mixed signal plot.
* On right side - Different WTA Networks.



Comparative Model Analysis

RAVSim presents an advanced feature, allowing users to conduct a comprehensive comparative analysis of various SNN models. This feature significantly enhances the understanding of model performance and provides assistance in the selection of the most suitable model for specific SNN applications.
The user is given the flexibility to make a choice regarding the type of model comparison they wish to perform. The options available are as follows:
* One-to-One Model Comparison - By selecting this option, the user can conduct a detailed comparative analysis between two specific neural models. This mode is particularly valuable when pinpointing subtle differences and performance variations between individual two models is essential.
* All Models Comparison - Alternatively, the user can opt for the "All Models Comparison" mode, which streamlines the process of comparing all available neural models simultaneously. This feature is especially useful when seeking an overall assessment of model performance across the entire set, making it an efficient choice for broader evaluations and quick insights.


input
input
input
input

* Figure (A) - All Model Performance and Accuracy Comparison Visualization VI
* Figure (B) - Visualization and Result VI for Spiking Activity Comparison of All Models
* Figure (C) - Visualization and Result VI for Computational Complexity Comparison of All Models
* Figure (D) - Visualization and Result VI for Biological Plausibility Comparison of All Models




Image Classification Task

This virtual instrument serves as the gateway to the image classification simulation environment. Within this environment, users have the capability to input their preferred parameter values, crucial for the analysis and visualization of image classification processes. This functionality allows users to explore and experiment with different datasets.
Upon completion of the analysis, the system will present the user with a comprehensive set of results and details for their review. These details include:
* Model Accuracy: This is visualized through plot representations, providing an in-depth view of how well the model performs in the task.
* Testing Error Rate: The analysis also reveals the testing error rate using MSE. This metric offers the accuracy of the model's predictions.
* Model Execution Time: The time it takes for the model to complete the image classification task is also reported.
* Dataset Details: Users can access detailed information about the dataset used for the analysis.

input

* On left side - Image classification VI after analysis
* On right side - RAVSim the weight generation feature.



input


Preprocess Custom Dataset Feature

Dataset preprocessing VI, offering users an adaptable platform for preprocessing custom datasets. Within dataset preprocess VI, users are allowed to use downloaded images for creating a scalable dataset. This flexibility includes the ability to use images with varying extensions, sizes, and qualities, pixel to the user's specific requirements.
* The tool supports a wide variety of image formats, accommodating datasets that suit diverse research and application needs.
* RAVSim also includes a friendly feature that enables users to preview images both before and after creating a dataset.
* This feature allow users can visually assess the content and quality of the images without the need to open each one individually.



Standard Deviation Error Rate

The standard deviation between each runtime is crucial because it allows end-users to detect invalid input values that may cause abnormal simulation times. The runtime standard deviation calculated from 10 simulation runs using different set of neurons.

input

* Time taken along with an error bar expressing standard deviation
* sing the image classification algorithm analysis for ten different test cases.



Citation

@article{sanaullah2023exploring,
  title={Exploring spiking neural networks: a comprehensive analysis of mathematical models and applications},
  author={Sanaullah and Koravuna, Shamini and R{\"u}ckert, Ulrich and Jungeblut, Thorsten and others},
  journal={Frontiers in Computational Neuroscience},
  volume={17},
  year={2023},
  publisher={Frontiers Media SA}
}
@article{sanaullah2023evaluation,
  title={Evaluation of Spiking Neural Nets-Based Image Classification Using the Runtime Simulator RAVSim.},
  author={Sanaullah and Koravuna, Shamini and R{\"u}ckert, Ulrich and Jungeblut, Thorsten and others},
  journal={International Journal of Neural Systems},
  pages={2350044--2350044},
  year={2023}
}
@inproceedings{sanaullah2022snns,
  title={SNNs model analyzing and visualizing experimentation using RAVSim},
  author={Sanaullah and Koravuna, Shamini and R{\"u}ckert, Ulrich and Jungeblut, Thorsten},
  booktitle={International conference on engineering applications of neural networks},
  pages={40--51},
  year={2022},
  organization={Springer}
}