Table 1.
νbias | Mean firing rate of bias Poisson spike train | All figures | 1000Hz |
σ | Noise amplitude | All figures, except Figure 1 | 3· 10−11 A/s0.5 |
Figure 1 (left) | 2· 10−11 A/s0.5 | ||
Figure 1 (right) | 3· 10−10 A/s0.5 | ||
Figure 1 (bottom) | 1· 10−9 A/s0.5 | ||
β | Exponential factor (fit) | All figures | 2.044· 109A−1 |
γ | Baseline firing rate (fit) | All figures | 8808Hz |
τr | Refractory period | All figures | 4ms |
τsyn | Time constant of recurrent, and bias synapses | All figures | 4ms |
τbr | “Burn-in” time of the neural sampling | All figures | 10ms |
gL | Leak conductance | All figures | 1nS |
urst | Reset potential | All figures | 0V |
C | Membrane capacitance | All figures | 10−12F |
θ | Firing threshold | All figures | 100mV |
W | RBM weight matrix (ϵ ℝNυ × Nh) | Figure 4 | N(−0.75, 1.5) |
bυ, bh | RBM bias for layer υ and h | Figure 4 | N(−1.5, 0.5) |
Nυ, Nh | Number of visible and hidden units in the RBM | Figure 4 | 5,5 |
Figures 7, 8, 7 | 824, 500 | ||
Nc | Number of class label units | Figures 7, 8, 7 | 40 |
2T | Epoch duration | Figures 4, 7, 8 | 100ms |
Figure 9 | 300ms | ||
Tsim | Simulation time | Figure 2 | 5s |
Figure 4 | 1000s | ||
Figure 7 | 0.2s | ||
Figure 9 | 0.85s | ||
Figure 8 (testing) | 1.0s | ||
Figure 8 (learning) | 2000s | ||
τSTDP | Learning time window | Figure 7 | 4ms |
η | Learning rate | Standard CD | 0.1· 10−2 |
Event-driven CD | 3.2· 10−2 |
Software simulation scripts are available online (https://github.com/eneftci/eCD.