P318 Algorithmic solutions for spike-timing dependent plasticity in large-scale network simulations with long axonal delays
Jan N. Vogelsang*1,2, Abigail Morrison*2,3, Susanne Kunkel1
1 Neuromorphic Software Ecosystems (PGI-15), Jülich Research Centre, Jülich, Germany 2RWTH Aachen University, Aachen, Germany 3Institute for Advanced Simulation (IAS-6), Jülich Research Centre, Jülich, Germany
*Email: j.vogelsang@fz-juelich.de Introduction
The precise timing of neuronal communication is a cornerstone in understanding learning and synaptic plasticity. Spike-timing dependent plasticity (STPD) models in particular rely on the precise temporal difference between pre- and post-synaptic spikes to adjust synaptic strength, where both the diverse axonal propagation delays and dendritic backpropagation delays play a crucial role in determining the precise timing. However, neural simulators, such as NEST, have traditionally represented transmission delays between neurons as a single aggregate delay value because of algorithmic challenges. We present two simulation frameworks addressing these challenges and validate across a set of small- to large-scale benchmarks.
Methods The NEST simulator reference implementation currently treats the entire delay as dendritic, which allows performing synaptic strength adjustments immediately after the occurrence of a pre-synaptic spike, avoiding costly buffering of spikes. This is an acceptable approximation for small networks but leads to inaccuracies when modeling long-range connections. In this framework, introducing axonal delays causes causality issues. At the time a pre-synaptic spike occurs, post-synaptic spikes only occurring in future time steps might reach the synapse before such spike due to predominant axonal delays. In order to mitigate this issue, one must either correct the weight on later occurrence of such post-synaptic spikes or postpone the STDP update. Results Both approaches were implemented and rigorously benchmarked in terms of runtime efficiency and memory footprint for varying synaptic delays and delay partitions. Correcting faulty synaptic updates achieves exceptional performance for fractions of axonal delay equal to or lower than the corresponding dendritic one. Only in the case of predominant and long axonal delays it starts to be outperformed by the alternative approach, which however required fundamental changes to the simulation framework to enable efficient buffering of individual spikes at the synapse level. However, benchmarks show that the approach induces a negative impact on performance for simulations not involving STDP dynamics, unlike the correction-based approach. Discussion Although different axonal and dendritic contributions are known to bias the synaptic drift towards either systematic potentiation or depression, there is a lack of simulation studies investigating the effects on network dynamics and learning in large neuronal systems. The ability to differentiate between axonal and dendritic delays represents a significant advance in neural simulation technology, as it addresses a long-standing limitation in spike-timing dependent plasticity modeling in large-scale, distributed simulations and enables future research in learning and plasticity, in particular, investigations of brain-scale models with STDP faithfully representing heterogeneous long axonal delays between areas.
Acknowledgements I want to thank Dennis Terhorst and Anno Kurth for assistance in benchmarking and running all the required jobs on the HPC systems. References -
