Paper
BioLCNet: Reward-modulated Locally Connected Spiking Neural Networks
Hafez Ghaemi, Erfan Mirzaei, Mahbod Nouri, Saeed Reza Kheradpisheh
DOI: https://doi.org/10.1007/978-3-031-25891-6_42
arXiv: https://arxiv.org/abs/2109.05539
Requirements
To install bindsnet you should only use the following command:
!pip install -q git+https://github.com/bindsnet/bindsnet
This will downgrade some of your installed packages including PyTorch.
Main Experiments
Feature Extraction
In this part, we trained our hidden layer to extract features from the MNIST images. We use these features as pre-trained weights in the classification task.
Image Classification
After transfering the weights from the pretrained features from the previous section, we train the network to classify the MNIST dataset images. You can also run "main.ipynb" for this experiment.
Classical(Pavlovian) Conditioning
In this experiment, we present the network with images belonging to one arbitrary class of the MNIST dataset as the neutral stimuli, and give reward such that it becomes conditioned to the desired response during each task. The purpose of this experiment is to show the effectiveness of our rewarding mechanism.