Table 3.
Simulation parameters for training the convolutional layers in ReStoCNet.
| Parameters | Values | ||
|---|---|---|---|
| C1 | C1 | C2/C3 | |
| Input dataset | MNIST | CIFAR-10 | CIFAR-10 |
| Maximum synaptic weight (whigh) | +1.0 | +1.0 | +1.0 |
| Minimum synaptic weight (wlow) | −1.0 | −1.0 | −1.0 |
| Weight initialization constant (αweight_init) | 75 | 30 | 30 |
| Simulation time-step, Δtsim | 1 ms | 1 ms | 1 ms |
| Simulation period for STDP, TSTDP | 25 ms | 25 ms | 25 ms |
| Maximum input spike rate for STDP | 200 Hz | 200 Hz | 500 Hz |
| Dropout probability for STDP, pdrop | 0.5 | 0.5 | 0.5 |
| STDPstride | 5 | 5 | 5 |
| Pre-trace decay time constant, τpre | 1.45 ms | 1.45 ms | 1.45 ms |
| preHebb_pot (eHB-STDP) | 0.50e-1 | 0.20e-1 | 0.20e-1 |
| preantiHebb_dep (eHB-STDP) | 0.50e-2 | 0.50e-2 | 0.50e-2 |
| pHebb_pot (eHB-STDP) | 0.01 | 0.05 | 0.05/25 |
| pantiHebb_dep (eHB-STDP) | 0.01 | 0.01 | 0.01/25 |
| pHebb_dep (eHB-STDP) | 0 | 0 | 0 |
| preHebb_dep (iHB-STDP) | – | 0.20e-1 | 0.20e-1 |
| preantiHebb_pot (iHB-STDP) | – | 0.50e-2 | 0.50e-2 |
| pHebb_dep (iHB-STDP) | – | 0.05 | 0.05/25 |
| pantiHebb_pot (iHB-STDP) | – | 0.01 | 0.01/25 |
| pHebb_pot (iHB-STDP) | – | 0 | 0 |
| Leaky-Integrate-and-Fire (LIF) neuron leak time constant, τmem | 9.5 ms | 9.5 ms | 9.5 ms |
| Rate of increase of LIF neuronal firing threshold, βthresh | 6e-4 | 6e-4 | 6e-4 (C2) |
| 8e-4 (C3) | |||
| Integrate-and-Fire (IF) neuron pooling threshold, θpool | 0.80 | 0.80 | 0.80 |
| Simulation period to estimate spiking activation, Tsim | 100 ms | 100 ms | 100 ms |
| Maximum input spike rate to estimate spiking activation | 500Hz | 500Hz | 500Hz |
| Low-pass filter time constant to estimate spiking activation, τlpf | 99.5 ms | 99.5 ms | 99.5 ms |