guest ::
| Home > Publications database > New biophysical mechanisms in NEST |
| Home > Publications database > New biophysical mechanisms in NEST |
| Talk (non-conference) (Invited) | FZJ-2025-03439 |
2024
Abstract: While the focus of this workshop is on the GPU kernel of NEST in preparation of exascale computers like JUPITER, the extension of the CPU kernel of NEST by further biophysical mechanisms is continuing. In this way, researchers immediately profit from advanced features and the functionality of the CPU code serves as a reference for the GPU implementation. The talk discusses two mechanisms recently added to the CPU code. The first is a framework for neuron-astrocyte interactions. There is presently no consensus on the role of astrocytes in the dynamics and plasticity of neuronal networks. In addition, the equations are complex because the relevance of individual details is unknown. Nevertheless, it is evident now that astrocytes play a role in plasticity, cognition, and behavior. Therefore, the attention of neuroscience to astrocytes is growing and research in the area is expanding. Computational work has so far been restricted to small networks due to the long observation times required and the lack of suitable simulation code. As an example, we discuss a network where astrocytes deliver slow calcium-governed currents (SIC) to postsynaptic neurons. The network exhibits oscillations controlled by neuron-astrocyte interaction. Astrocytes require a generalization of the connectivity concepts of NEST from pairwise rules to tripartite motifs. The framework provides an implementation compatible with the usual hybrid parallelization of the CPU code.The second is a framework for backpropagation-like learning for spiking neuronal networks based on only local information. The theory for this, called eligibility propagation (e-prop), was published by Bellec et al. (2020) together with a time-driven algorithm for TensorFlow. The talk explains the reformulation of the algorithm for the event-driven update of synapses in the NEST code and demonstrates the reproduction of original test cases. This naturally leads to further modifications like the asynchronous update of synaptic weights that maintain learning performance while enhancing biological plausibility. In this way constraints of the original theory are relaxed, and the learning scheme is available for large-scale spiking network models.
|
The record appears in these collections: |