KIP-Veröffentlichungen

We present analog accelerated neuromorphic hardware, specifically BrainScaleS-2, as a reliable and efficient tool for studying high-dimensional biologically-plausible synaptic plasticity. We focus on a calcium-based synaptic plasticity rule that integrates different neuron and synapse variables on multiple timescales. Compared to our previous work, we automate the calibration of the analog circuits which emulate the calcium dynamics and scale the plasticity rule to multiple synapses using the SIMD extension of the embedded digital processor. To validate our implementation, we used standard stimulation protocols. The analog nature of the hardware, reduced-precision arithmetic, and the larger integration time steps of the plasticity rule update compared to the simulation cause numerical differences between the emulation and simulation results. Nevertheless, our results show that the implementation can faithfully emulate the model dynamics for multiple synapses in parallel. The final judgment on the implementation of the plasticity rule depends on its applications, specifically modeling cognitive functions in computational neuroscience and machine learning. Our approach aims at exploiting the energy efficiency and acceleration factor of BrainScaleS-2 as well as its analog nature for the study and application of synaptic plasticity.