Dataset Creation

Convert Images to Poissonian Spikes

Convert your image data to a Poisson spike source to be able to use with Spiking Neural Networks.

Technique

Poisson Spike Train Generation

Rate Coding We used the Poisson pixel intensity technique to encode visual stimuli into neural spike trains. This method translates the brightness of each pixel into the probability of generating a spike, simulating the rate coding observed in biological neurons. By applying this technique to images with different line orientations and positions, we created spike trains that accurately represent the visual input, enabling our SCNN model to replicate the orientation-selective behavior of V1 neurons.

→

→

→

The parameters below are used when running convert_image_to_spike_array.py in order to turn pumpkins above into a spike array.
max_freq = 60000 (Hz)
on_duration = 10000 (ms)
off_duration = 5000 (ms)

Model Implementation

The feedforward multilayer spiking convolutional neural network (SCNN) model has an input layer realizing the contrast cell response. The two hidden layers realize the simple cell response and the complex cell response. The output layer consists of a single output neuron. The complex cell response is subject to the Bienenstock-Cooper-Munro (BCM) learning rule. The Leaky Integrate-and-Fire (LIF) Neurons are part of all layers. The machine learning algorithm is unsupervised.

Results & Analysis

Conducting a Two-Way ANOVA (type III) per simple cell kernel on the two factors "orientation" (levels: 0,45,90,135) and "position" (levels: 0,1,2,3). Vizualising the boxplots accordingly. Second visualization expands the boxplot by information about the evolution of the ouput during training. Make sure you downloaded the file ouput_data first. To access the obtained plots see NeuroNetV1/output_data /output_data.txt.

Running the code

To specify the model architecture update model.py. This is necessary when specifying the simple cell kernel in the function simple_cell_kernel(gamma = 1). To make sure each of the four model architectures make use of the same generated data that is used for multiple iterations of your network model run generate_data() in train_model.py only once after four epochs.

Requirements

• matplotlib (3.0.3)
• numpy (1.17.3)
• opencv-python (4.1.1.26) Run pip install -r requirements.txt to install them all.

Project Files and Their Usage

dataset_generation/
  ├── convert_image_to_spike_array.py
  ├── data_images_new
  ├── raster_plots_new
  ├── txt_files_new
  └── draw_image.py
model/
  ├── model.py
  ├── train_model.py 
analysis/
  ├── analysis.py
output_data/
  ├── output_data.txt
├──requirments.txt

convert_image_to_spike_array.py is the main file.

• Please see its usage by running it: python convert_image_to_spike_array.py images 1000 200 100
• The program will store the output spike array as a txt file under txt_files_new/_ folder in the same directory after the run.
• The program will create a folder called raster_plots_new, tha twill store raster plots with spike trains for each image.
• You may use a single image file (extension could be anything OpenCV accepts) or a folder which contains multiple images (extensions need to be .png) as input.

draw_image.py enables you to draw your own images by adding simple shapes into it via OpenCV. For more information please see the file.

model.py specifies the model architecture

train_model.py allows you to train the model and initialize the weights

analysis.py After runing the model (or downloading ouput_data) the script provides the analysis and respective visualizations