界面已本地化;技能介绍与说明文档按所选语言由 AI 自动翻译,代码、命令、路径和专业术语保留原文,以原作者文档为准。
探索技能
效率与协作 · SKILL.md

synapses-and-inputs

synapses-and-inputs

【效率与协作】Define Brian2 2.9.0 synaptic connectivity, event pathways, delays, plasticity, and explicit or stochastic input sources.

独立 GitHub 快照 · 自动分类 · 以原作者文档为准

技能说明

GitHub

name: synapses-and-inputs description: "Define Brian2 2.9.0 synaptic connectivity, event pathways, delays, plasticity, and explicit or stochastic input sources." disable-model-invocation: true metadata: disco-role: operating license: NOASSERTION

Synapses and inputs

Use this route when the task connects spike-producing groups, applies synaptic code, configures delays or plasticity, or supplies Poisson, timed, or explicit spike input in Brian2 2.9.0. The route is deliberately workflow-oriented: first define the source/target and synaptic equations, then create connections, assign existing synapse state, and only then run and inspect event effects.

Route map

  • API signatures, indexing, pathway names, and input-class boundaries: api-reference.md
  • End-to-end construction, timing, replay, and validation recipes: workflows.md
  • Conditional/probabilistic connectivity, event-driven traces, STDP, and summed variables: connectivity-and-plasticity.md
  • Install/import, optional dependency, data/configuration, API misuse, timing, and workflow recovery: troubleshooting.md
  • Run a tiny deterministic connectivity and delayed-spike assertion: connectivity_smoke.py

Operating contract

  1. Keep neuron equation semantics, units, thresholds, resets, and state-updater choice with modeling and units-and-equations. This route consumes a valid spike source and target variable.
  2. A Synapses constructor defines model/pathways; it does not create synapses. Call connect(...) before indexing or assigning synaptic arrays.
  3. Use explicit _pre/_post suffixes when a name could be ambiguous. Assign w, delay, and other synapse state after the relevant connections exist.
  4. Treat event-driven variables as values updated at event times, not as continuously observable traces. Ensure their dependencies remain compatible with event-driven integration.
  5. Choose PoissonGroup when individual spikes matter, PoissonInput when only aggregate independent input matters, SpikeGeneratorGroup for reproducible event schedules, and TimedArray for sampled time functions.
  6. Validate units, dt/delay alignment, event binning, and output timing with a small fixture before scaling. Monitor analysis belongs to recording; device and standalone restrictions belong to code-generation.

The bundled smoke script is read-only with respect to the repository and uses only a tiny in-memory runtime simulation; it does not download data or run native repository tests.

我的收藏