2.5. Software Simulation With NEST

In order to compare the learning performance and speed of the chip to a network of deterministic, perfectly identical LIF neurons, we ran a software simulation of the experiment using the NEST v2.14.0 SNN simulator (Peyser et al., 2017), using the same target LIF parameters as in our experiments using the chip, with time constants scaled by a factor of 103. We did not include fixed-pattern noise of neuron parameters in the simulation, i.e., all neurons had identical parameters. We used the iaf_psc_exp integrate-and-fire neuron model available in NEST, with exponential PSC kernels and current-based synapses. Using NEST's noise_generator, we are able to investigate the effect of injecting Gaussian current noise into each neuron. The scaling factors η₊ and η₋, as well as the time constants τ₊ and τ₋ of the correlation sensors were chosen to match the mean values on BSS2. The correlation factor a₊ was calculated within the Python script controlling the experiment using Equation (1) and the spike times provided by the NEST simulator. Hyperparameters such as learning rate and game dynamics (e.g., the reward window defined in Equation 5) were set to be equivalent to BSS2 and weights were scaled to a dimensionless quantity and discretized to match the neuromorphic emulation.

The synaptic weight updates in each iteration were restricted to those synapses which transmitted spikes, i.e., the synapses from the active input unit to all output units (32 out of the 1, 024 synapses), as the correlation a₊ of all other synapses is zero in a perfect simulation without fixed-pattern noise. This has the effect of reducing the overall time required to simulate one iteration and is in contrast to the implementation on BSS2, where all synapses are updated in each iteration as there is no guarantee that correlation traces are zero and we excluded this kind of “expert knowledge” from the implementation.

The source code of the simulation is publicly available (Wunderlich, 2019).