Get started

Overview

Learned neuron models for fast population simulation and inference-time adaptation.

AxoSim approximates the response of detailed biological neurons to synaptic input histories. You can load trained GRU and Mamba models, construct populations with explicit dendritic contacts, and adapt neuronal and synaptic parameters through a complete temporal sequence.

AxoSim-GRU Medium and Mamba Medium in color, alongside Branch-ELM and TCN models and the CoreNEURON reference in gray, compared on inference and history throughput, dataset-relative voltage and subthreshold fidelity, and spike Mean F1. AxoSim-GRU Medium and Mamba Medium in color, alongside Branch-ELM and TCN models and the CoreNEURON reference in gray, compared on inference and history throughput, dataset-relative voltage and subthreshold fidelity, and spike Mean F1.

The axes show (i) inference throughput, counting 1-ms neuron predictions per second; (ii) voltage fidelity, the voltage SERA skill as a fraction of the skill of the teacher voltage clipped below action potentials; (iii) subthreshold fidelity, the RMSE skill of the voltage outside action-potential windows; (iv) spike Mean F1, averaged over 0–5 ms matching tolerances; and (v) history throughput, counting input milliseconds processed per second for one supplied 2-second history. The fidelity axes are dataset-relative: each learned model is scored on held-out traces of its own training data, AxoBench for AxoSim and NeuronIO for Branch-ELM and TCN, with firing rates matched between the datasets. CoreNEURON is the ground-truth detailed simulator; Branch-ELM, the learned-surrogate predecessor of AxoSim, and TCN are released reference models. AxoSim–GRU Medium and Mamba Medium are compared with three Branch-ELM sizes and the smallest and largest TCN; larger radii indicate better performance. Raw measurements, fixed mappings and runtime contracts appear in the appendix of the technical report.

Start with a neuron model

Use model selection to choose a temporal architecture and backend, then load a checkpoint to predict spike logits and somatic voltage. The common single-neuron contract is an input tensor shaped (batch, time, input channels) and an output tensor shaped (batch, time, 2).

The trained architectures extend the branching mechanism of Branch-ELM. Signed input events preserve their dendritic destinations, while GRU or Mamba temporal cores model the response history. Use history processing when complete input histories are available, or streaming inference when inputs arrive over time.

Connect and adapt populations

The population interface combines a compact AxoSim-Lite neuron model with morphology identities, dendritic contact routes, and independently mutable synaptic efficacies. Sparse events preserve contact identity without allocating a dense contact history. Inference-time adaptation selects behavior, morphology, and synaptic parameter groups while differentiating through the supplied sequence.

The optimized runtime evaluated in the report simulates 65,536 neurons with 4,000 retained contacts per neuron at 3.38× real time on an RTX 5090. Full BPTT reaches real time for 20,000 neurons with 2,000 retained contacts. These measurements describe the frozen benchmark implementation. Use create_population for connected rollout, or AxoSimPopulation to fit supplied Lite histories. Reproducibility and deployment explains how to recover the reported evidence and which execution paths are covered by the measurements.

Benchmark with AxoBench

AxoBench provides the datasets, metrics, and evaluation protocol for single-neuron fidelity. It accepts any NumPy prediction function and depends only on NumPy, Matplotlib, and Requests.

  • Datasets. Detailed-teacher simulations in three releases: single-neuron traces from five layer-5 pyramidal morphologies, population contexts whose inputs were recorded in connected AxoSim networks, and paired interventions that change one input sample in a defined way. NeuronIO simulations convert to the same format.
  • Metrics. The core metrics are Voltage SERA, spike Mean F1 over 0–5 ms tolerances, the subthreshold voltage RMSE, and Response SERA. Skill rows express SERA against a constant voltage and against the best prediction without action potentials; their ratio and a firing-rate reweighting support comparisons across datasets.
  • Evaluation. The evaluator calibrates the spike threshold on traces that are not scored and compares a candidate with a cached baseline on identical traces, with paired bootstrap intervals.
  • Diagnostics. Every report ends with a fixed 24-row attention panel. The error anatomy breaks a model’s errors down by spike outcome, time from a spike, firing rate, frequency band, and morphology. Controlled defects test how each metric responds, and diagnostic rows export to JSON, JSONL, or CSV.
  • Plots. axobench diagnostic-plots and axobench error-anatomy write paper-style figures as SVG, PDF, or PNG; the error-anatomy figure is also exported as a standalone pgfplots file for LaTeX.

Start with choose a dataset, then benchmark a model.

Follow the lifecycle

  1. Install AxoSim and inspect the model presets.
  2. Choose a model, run inference, and train on your data.
  3. Build a population, then use sparse contact events for sparse inputs.
  4. Adapt parameter groups and capture repeated CUDA updates when shapes are fixed.
  5. Choose an AxoBench dataset, evaluate the checkpoint with the AxoBench protocol, and preserve reproducibility records.
  6. Use the AxoSim API reference and the AxoBench API reference for exact signatures, defaults, and source links.

Cite AxoSim

The technical report is by Davide Wiest and Jonathan Schäfer, Axym Labs. Its central figure and dataset-relative comparison use AxoBench 0.2 metrics: SERA per teacher spike, the subthreshold voltage error, and spike Mean F1. Its ordinary-AxoBench results still use the earlier SERA definition and Dynamics SERA, so those values are not comparable with AxoBench 0.2’s; see core metrics.

@techreport{wiest2026axosim,
  title = {AxoSim: Learned Neuron Models for Population Simulation at Scale},
  author = {Wiest, Davide and Schäfer, Jonathan},
  institution = {Axym Labs},
  year = {2026}
}

Search guides, examples, and API signatures.