Canonical Parameters
Canonical network and training parameters for COBA and PING behaviour.
These are canonical parameters for decent conductance-based (COBA) and pyramidal–interneuron gamma (PING) behaviour, following the shared TR01/TR02 recipe in exp022. E and I denote excitatory and inhibitory neurons; paired values below are E / I. LIF means leaky integrate-and-fire.
| Parameter | Canonical value |
|---|---|
| Network and input | |
| Input / hidden E / hidden I / output neurons | 784 / 1,024 / 256 / 10 |
| Input encoding | Independent Poisson pixel channels; 25 Hz at maximum pixel brightness |
| Presentation / timestep | 200 / 0.1 ms; 2,000 updates per presentation |
| State between presentations | Reset |
| E→I→E loop strength | COBA: 0; PING: 1 |
| E→E / I→I recurrence | Absent |
| Recurrent connection delay | One timestep (0.1 ms) |
| Neuron and synapse dynamics | |
| Hidden neurons | Conductance-based LIF |
| Membrane capacitance, E / I | 1.0 / 0.5 nF |
| Leak conductance, E / I | 0.05 / 0.10 µS |
| Membrane time constant, E / I | 20 / 5 ms (capacitance divided by leak conductance) |
| Resting / reset / threshold voltage | −65 / −65 / −50 mV |
| Refractory hold, E / I | 1.2 / 0.6 ms |
| AMPA / GABA synaptic decay | 2 / 6 ms |
| Readout | Ten spiking LIF neurons; temporal mean of pre-reset voltage supplies class logits |
| Readout time constant / soft-reset threshold | 2 ms / 1 (dimensionless threshold) |
| Weights and learning | |
| Input-weight initialization | Normal parent mean 0.9, standard deviation 0.09; clamp negatives to zero, randomly zero 95%, then divide by (0.05 × 784) |
| PING E→I initialization | Normal mean 1, standard deviation 0.1; clamp negatives to zero and divide by 1,024 |
| PING I→E initialization | Normal mean 2, standard deviation 0.2; clamp negatives to zero and divide by 256 |
| COBA recurrent weights | Zero |
| Readout-weight initialization | Normal mean 1.12060546875, standard deviation 0.8349609375; clamp negatives to zero; no further fan-in normalization |
| Trainable projections | Input→E and E→readout; recurrent weights fixed. Initial input and readout zeros can regrow |
| Weight constraints / readout bias | Nonnegative weights, inhibitory action through I→E; no readout bias |
| Optimizer / learning rate | AdamW / 0.0004 |
| Weight decay / gradient-norm clip | 0 / 1 |
| Voltage-gradient damping / surrogate slope | 1,000 / 1 for both COBA and PING |
| Adaptive threshold / learned leak / state clamping | Disabled |
| TR01 and TR02 training | |
| Dataset | MNIST; official test partition excluded from optimization and checkpoint selection |
| Epochs / minibatch / seeds | 50 / 256 / 42, 43, 44 |
| TR01 training / validation samples | 54,000 / 6,000; 60,000-sample pool |
| TR01 activity penalty | Off |
| TR02 training / validation samples | 6,300 / 700; 7,000-sample pool |
| TR02 hidden-E activity ceiling | Off, 25, 10, 5, 2.5 or 1 Hz |
| TR02 penalty when enabled | 0.041 × mean squared excess above the ceiling; compute each presentation’s population-mean E rate in Hz before applying the penalty |
Source: exp022’s recipe and network and weight initialization and training loop.