The architecture of this project uses the CSNN proposed in STDP-based spiking deep convolutional neural networks for object recognition
The coding library uses the Spyketorch. The Github Page of SpykeTorch can be found here
This porject implements the Convolutional Spiking Neural Network (CSNN) with SpykeTorch. We test the performance of CSNN on the MNIST dataset, and compare the results with Temporal Neural Network (TNN). We implement the Band Corelator and General Corelator in the TNN modules to show that CSNN has a better performance compared to TNN. However, we did not achieve as high accuracy as mentioned in the paper, STDP-based spiking deep convolutional neural networks for object recognition. And we also find that traditional machine learning methods may provide better accuracy on MNIST compared to CSNN. Another improvement we made is that we use more classifiers rather than only using SVM. We found that a neural network with only one hidden layer will perform better than using SVM as the final classifier.
- Python 3.6
- Numpy
- Pytorch
- SpykeTorch
In our implementation, we add two more filters: ON-Center/OFF-Center filters. If you want to use these two filters, please add the following codes into SpykeTorch/utils.py
class OnCenter(FilterKernel):
def __init__(self, window_size):
super(OnCenter, self).__init__(window_size)
def __call__(self):
kernel = np.zeros((self.window_size, self.window_size))
kernel /= -8.0
kernel[1][1] = 1.0
OC_tensor = torch.from_numpy(kernel)
return OC_tensor.float()
class OffCenter(FilterKernel):
def __init__(self, window_size):
super(OffCenter, self).__init__(window_size)
def __call__(self):
kernel = np.zeros((self.window_size, self.window_size))
kernel /= 8.0
kernel[1][1] = -1.0
OF_tensor = torch.from_numpy(kernel)
return OF_tensor.float()
python project/CTNN.py