causalformer

Implementation and results of Causalformer from Attention for Causal Relationship Discovery from Biological Neural Dynamics (accepted to the NeurIPS 2023 Workshop on Causal Representation Learning). Implementation of Causalformer is based on the Spacetimeformer. Please refer to the Spacetimeformer repo for dependencies of Causalformer.

Notes:

• izhikevich_data (directory): Simulated data used in the paper. Consists of 160 different random networks in total.
  • v_alltimes.txt: Variable $v$ of each neuron at each timestep. Size: $n \times T$, $n$: number of neurons, $T$: number of timesteps.
  • u_alltimes.txt: Variable $u$ of each neuron at each timestep. Size: $n \times T$.
  • v_normed_alltimes.txt: Normalized variable $v$ of each neuron at each timestep. For each neuron, $v$ is first clipped at the firing threshold and then z-scored by its mean and std across timesteps. Size: $n \times T$.
  • inp.txt: Random thalamic input to each neuron at each timestep. Size: $n \times T$.
  • firings.txt: 1st column: timestep at which a spike is generated. 2nd column: the neuron who fired the spike at the correpsonding timestep in the 1st column.
  • params_abcd.txt: Parameters $a, b, c, d$ for each neuron. Size: $n \times 4$.
  • connectivity.txt: Connectivity matrix $M$. $M(i, j) > 0$: excitatory connection from neuron $j$ to $i$. $M(i, j) < 0$: inhibitory connection from neuron $j$ to $i$. $M(i, j) = 0$: no connection from neuron $j$ to $i$. Magnitude of $M(i, j)$ indicates connectivity strength. Size: $n \times n$.
  • mvgc_{z}_F_p{x}.txt: F-test statistic for time-domain pairwise conditional Granger causalities, returned by the MVGC routine. {x}: MVGC maximum model orders. {z}: 'aic' or 'bic', model selection criterion.
• run_izhikevich_sim.m: Matlab code for data simulation using the Izhikevich model and running Multivariate Granger Causality (MVGC) analysis.
• causalformer_model (directory): Pytorch implementation of Causalformer.
• randnet_results (directory): Summary of experiment results on the data from izhikevich_data.
• plot_randnet_results.ipynb: Notebook for visualizing results in randnet_results. Reproduces figure 2, figure 4(b) in the paper.
• example_trained_models (directory): Trained Causalformer models on one example random network from izhikevich_data (5 neurons, connection probablity 0.4, random topology 4). Consists of 10 models with different random initializations (i.e., random seeds).
• results_visualization.ipynb: Notebook for visualizing results in example_trained_models. Reproduces figure 1(a)(b)(d)(e), figure 4(a) in the paper.