Table 1 Hyperparameters for training.
From: Biologically-informed excitatory and inhibitory ratio for robust spiking neural network training
Parameter | Value |
|---|---|
Membrane time constant \(\tau _{mem}\) | 10 ms |
Synaptic time constant \(\tau _{syn}\) | 5 ms |
Simulation time step dt | 1 ms |
Input layer size \(L_{0}\) | 784 (FMNIST), 700 (SHD) |
Hidden layer size \(L_{1}\) | 100 (FMNIST), 200 (SHD) |
Output layer size \(L_{2}\) | 10 (FMNIST), 20 (SHD) |
Batch size | 256 |
Learning rate \(R_L\) | 0.01 |
Optimizer | Gradient descent |
Input to hidden initial weight \(\sigma _{L0,1}\) | 0.0005 to 0.075 (FMNIST), 0.00005 to 0.003 (SHD) |
Hidden to output initial weight \(\sigma _{L1,2}\) | \(7(1-e^{\frac{dt}{\tau _{mem}}})/\sqrt{L_1}\) (FMNIST), 0.0005 to 0.003 (SHD) |
Excitatory: inhibitory hidden layer ratio | 50:50 to 100:0 |
