Companion code to Confavreux*, Ramesh*, Gonçalves, Macke* & Vogels*, NeurIPS 2023, Meta-learning families of plasticity rules in recurrent spiking networks using simulation-based inference
https://openreview.net/forum?id=FLFasCFJNo

• Have git lfs installed before cloning the repo (see https://git-lfs.com/)
• Set-up the repo and dependencies with setup.py (see section below for more details).
• Install and test Auryn (see section below for more details).
• Compile the Auryn network simulations (see section below for more details).

Go through the notebooks:

• 1_Fit_posterior.ipynb
• 2_Sample_from_posterior.ipynb
• 3_Simulate_samples.ipynb
• 4_Compute_metrics.ipynb
• 5_Analysis_example.ipynb

• Create and activate a new conda virtual environment https://docs.conda.io/projects/miniconda/en/latest/miniconda-install.html
• Install python, pip and setuptools: conda install python
  Note: this may install a python version that is too recent for some of the dependencies (e.g. pytorch), in which case force install an earlier version: e.g. conda install python=3.11
• Install fSBI with setup.py: move to fSBI directory + python3 -m pip install -e .
• Install jupyter notebook and start the tutorial

All spiking network simulations in this repo use Auryn, a fast, C++ simulator developped by Friedemann Zenke. To install, please refer to https://fzenke.net/auryn/doku.php?id=start
Note that installing Auryn with MPI support is not required for the tutorial.

• Compile the auryn simulations sim_bg_IF_EEEIIEII_6pPol.cpp and sim_bg_CVAIF_EEIE_T4wvceciMLP.cpp located in synapsbi/simulator/cpp_simulators/. First, edit the Makefile in the same directory, you should only need to change AURYNPATH there.
  For troubleshooting, refer to https://fzenke.net/auryn/doku.php?id=manual:compileandrunaurynsimulations
• Go to tasks_configs/ and update auryn_sim_dir and workdir inside the 2 yaml files (these variables control where Auryn will write output spike trains).

The main data release alongside this paper are all the plasticity rules we simulated to obtain the posteriors for generally plausible plasticity rules (i.e. rules sampled from all posteriors along the filtering).

• data_synapsesbi/bg_IF_EEEIIEII_6pPol/bg_IF_EEEIIEII_6pPol_all.npy are all the rules from the polynomial search space (Fig 2).
• data_synapsesbi/bg_CVAIF_EEIE_T4wvceciMLP/bg_CVAIF_EEIE_T4wvceciMLP_all.npy are all the rules from the MLP search space (Fig 4).

These are numpy structured arrays containing the plasticity parameters and the network metrics computed after a simulation of a spiking network evolving with the rule in question.

Fields:

• theta: the values of the plasticity parameters.
  For the polynomial rules, this array is: [τ_{pre EE}, τ_{post EE}, α_EE, β_EE, γ_EE, κ_EE, τ_{pre EI}, etc... same for EI, IE and II].
  For the MLP rules: [η_EE, η_IE, W_{pre EE}, W_{post EE}, W_{pre IE}, W_{post IE},].
• seed: unique identifier for each rule
• metrics: rate, cv_isi, kl_isi', spatial_Fano, temporal_Fano, auto_cov, fft, w_blow, std_rate_temporal', std_rate_spatial, std_cv, w_creep, rate_i, weef, weif, wief, wiif. Refer to the paper for the definition of each metric