KIP-Veröffentlichungen

One application of neuromorphic computers is providing an energy-efficient solution for computer simulations of neural networks in the field of computational neuroscience. BrainScaleS-2 is a mixed-signal neuromorphic platform that promises accelerated and faithful emulation of biological neural networks. In this work, we aim to emulate the synaptic tagging and capture plasticity rule for a single synapse on BrainScaleS-2. The three main steps for emulating this rule are emulating the neural dynamics, emulating the calcium dynamics, and implementing the equations of the plasticity rule. The first two steps are achieved using the neuron and the adaptation circuits respectively of the analog core of BrainScaleS-2 that emulates the neuron dynamics based on the adaptive exponential leaky integrate-and-fire model. The third step relies on differential equations solved numerically on the digital plasticity processor of BrainScaleS-2. The main hardware constraints are the update timestep used to solve the differential equations and the use of finite arithmetic. Using a higher update timestep and stochastic rounding, we show that BrainScaleS-2 can faithfully emulate a single synapse that follows the synaptic tagging and capture plasticity rule for four stimulation protocols. The results show that almost no statistically significant results exist between the simulation and the emulation schemes, and that the variability of the test statistics mostly stems from the spikes. We conclude our work by reaffirming the role of BrainScaleS-2 in speeding-up the emulation of neural dynamics in computational neuroscience.