st |
Hidden external state at time t |
Section Details of Simulation |
rX,jt |
Firing rate of input neuron j at time t |
Equation (5) |
rY,it |
Firing rate of output neuron i at time t |
Equation (1) |
wij |
Synaptic weight from input neuron i to output neuron j |
Constant (Figures 1–3)Equation (2) (Figures 4–8) |
cij |
Number of connection from input neuron i to output neuron j (Note that here cij = 0 or 1) |
Section Synaptic Connection Learning |
ρij |
Connection probability from input neuron i to output neuron j |
Constant (Figures 1–4)Equation (3) (Figures 5, 6)Equation (4) (Figures 6I, 7, 8) |
|
The dual Hebbian rule |
Equation (2) + Equation (3) |
|
The approximated dual Hebbian rule |
Equation (2) + Equation (4) |
θjμ |
Response parameter of neuron j to hidden state μ |
Section Gaussian Model, Poisson Model |
qjμ |
Normalized response parameter of neuron j to hidden state μ. Especially in the Gaussian model, qjμ = θjμ/σX2 |
qjμ = h(θjμ) |
Ωμ |
Set of output neurons that selective for hidden state μ |
Section Accuracy of Estimation |
hw |
Input threshold |
Section Details of Simulation |
σX |
Noise in input neuron firing rate |
σX = 1.0 |
γ |
Parameter for sparseness of connectivity |
Sections Weight Coding and Connectivity Coding and Dual Coding and Cut-Off Coding |
bh |
Strength of homeostatic plasticity |
Equation (2) |
τc |
Timescale of rewiring |
Section Synaptic Connection Learning |
κm |
Ratio between constant and variable component in θjμ |
θjμ = 1Z[κmθjμconst+(1-κm)θjμvar] |
θconst, θvar |
Two component of input structures used in Figure 6 |
Section Gaussian Model |
T2 |
Interval between update of the variable component θvar |
T2 = 105 |
θctrl,θtraining |
Two input structures used for modeling control and training phases in Figure 8 |
Section Gaussian Model |