sunzhe839 / CRSNN

Unsupervised Conditional Reflex Learning Based on Convolutional Spiking Neural Network and Reward Modulation

Model

Structure

Requirements

1. NEST
2. Pytorch
3. SkypeTorch
4. V-REP (simulation platform)
5. ROS(interface between simulation environment and the script)

How to run

1. generate noise-like data with opencv and put the data into ./images
2. train the feature-extract unit with the generated noise-like data in an unsupervised manner:

   python ./feature_extract_unit/train_with_noiselikedata.py
   

3. train the decision-making unit:

   python ./decision_making_unit/train.py  
   

   • parameters of the network can be changed in ./decision_making_unit/parameters.py

   • note: ./decision_making_unit/environment_1.py corresponds to scenario 1~3 and ./decision_making_unit/environment_2.py corresponds to scenario 4.

Some Results

A. Visualization of the feature-extraction unit

B. Visualization of the decision-making unit

C. Training period

D. Testing performance

E. Generalization evaluation

F. Robustness evaluation

Acknowledge

Our work is based on the following works:

1. SkykeTorch
   [Paper | Project]
2. R-STDP
   [Paper | Project]