Trial matching: capturing variability with data-constrained spiking neural networks
Authors: Christos Sourmpis, Carl Petersen, Wulfram Gerstner, Guillaume Bellec
NeurIPS 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We focus specifically on the difficulty to match the trial-to-trial variability in the data. Our solution relies on optimal transport to define a distance between the distributions of generated and recorded trials. The technique is applied to artificial data and neural recordings covering six cortical areas. We find that the resulting RSNN can generate realistic cortical activity and predict jaw movements across the main modes of trial-to-trial variability. |
| Researcher Affiliation | Academia | Christos Sourmpis, Carl C.H. Petersen, Wulfram Gerstner, Guillaume Bellec Brain Mind Institute, School of Computer and Communication Sciences and School of Life Sciences, Ecole Polytechnique F ed erale de Lausanne (EPFL), CH-1015 Lausanne, Switzerland, {firstname.lastname}@epfl.ch |
| Pseudocode | No | The paper describes the model dynamics using mathematical equations but does not include any pseudocode or clearly labeled algorithm blocks. |
| Open Source Code | Yes | Our code is available at github.com/EPFL-LCN/pub-sourmpis2023-neurips (code DOI hosted by zenodo). |
| Open Datasets | Yes | This paper aims to model the large-scale electrophysiology recordings from [2], where they recorded 6,182 units from 12 areas across 18 mice 2. All animals in this dataset were trained to perform the whisker tactile detection task described in Figure 1: in 50% of the trials (the GO trials), a whisker is deflected and after a 1 s delay period an auditory cue indicates water availability if the mouse licks, whereas in the other 50% of trials (the No-Go trials), there is no whisker deflection and licking after the auditory cue is not rewarded. Throughout the paper we attempt to create a data-constrained model of the six areas that we considered to play a major role in this behavioral task: the primary and secondary whisker somatosensory cortices (w S1, w S2), whisker motor cortices (w M1, w M2), the primary tongue-jaw motor cortex (tj M1) and the anterior lateral motor cortex (ALM), also known as tj M2 (see Figure 1A and 3A). The reference [2] is "Vahid Esmaeili et al. Neuron 109.13 (2021), pp. 2183 2201.". |
| Dataset Splits | No | For the sound evaluation of the model, we separate the recorded trials of every session into a training set (75% of the total trials) and a testing set (25%), that is never used in the fitting procedure. The split is done in a stratified manner so both sets have the same distribution of trial types. The paper mentions training and testing sets, but no explicit validation set for hyperparameter tuning or early stopping. |
| Hardware Specification | Yes | The full optimization lasts for approximately one to three days on a GPU A100-SXM4-40GB. |
| Software Dependencies | No | The paper mentions software like 'pytorch' and 'scipy', but does not specify their version numbers or any other software dependencies with version information. |
| Experiment Setup | Yes | The spiking dynamics are then driven by the integration of the somatic currents It j,k into the membrane voltage vt j,k, by integrating LIF dynamics with a discrete time step δt = 2 ms. The jaw movement yt k is simulated with a leaky integrator driven by the activity of tj M1 and ALM neurons, followed by an exponential non-linearity. The membrane time constants m = 30 ms for excitatory and m = 10 ms for inhibitory neurons define j = exp and jaw = 50 ms define similarly jaw which controls the velocity of integration of the membrane voltage and the jaw movement. To implement a soft threshold crossing condition, the spikes inside the recurrent network are sampled with a Bernoulli distribution zt v0 )), where v0 is the temperature of the sigmoid (σ). The spike trains xt i model the thalamic inputs as simple Poisson neurons producing spikes randomly with a firing probability of 5 Hz and increasing their firing rate when a whisker or auditory stimulation is present (see Appendix A). The last noise source t j is an instantaneous Gaussian noise t j of standard deviation βvthr δt modeling random inputs from other areas (β is a model parameter that is kept constant over time). At each iteration until convergence, we simulate a batch of K = 150 statistically independent trials. |