Nonlinear multiregion neural dynamics with parametric impulse response communication channels
Authors: Matthew Dowling, Cristina Savin
ICLR 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We develop a state noise inversion free variational filtering and learning algorithm for our model and show, through neuroscientifically inspired numerical experiments, how the proposed model can reveal interpretable characterizations of the local computations within and the flow of information between neural populations. We further validate the efficacy of our approach using simultaneous population recordings from areas V1 and V2. |
| Researcher Affiliation | Academia | Matthew Dowling & Cristina Savin Center for Neural Science New York University EMAIL |
| Pseudocode | No | The paper includes mathematical equations and descriptions of algorithms (e.g., in Appendix A for 'STATE-NOISE INVERSION FREE FILTERING'), but no explicitly labeled 'Pseudocode' or 'Algorithm' blocks with structured steps. |
| Open Source Code | No | The paper does not contain any explicit statement about providing open-source code, nor does it provide a link to a code repository. |
| Open Datasets | Yes | Our final experiment asked whether our method can reveal neural correlates of cross-region communication from real neurophysiological recordings. To examine this, we considered simultaneous neural recordings taken from areas V1 and V2 of a macaque monkey as it observed gratings of different orientations on a screen (Zandvakili & Kohn, 2015). This dataset has been used in other studies to examine the efficacy of intraregional models of neural signaling (Gokcen et al., 2024; Li et al., 2024), making it a natural testbed for comparison. Like Li et al. (2024) and Gokcen et al. (2024), we used spiking activity from session 106r001p26. |
| Dataset Splits | Yes | Then, on a test set of 50 trials, we computed the MSE of a held-out set of neurons whose firing rate was inferred from a separate set of held-in neurons. Following a similar procedure as Li et al. (2024), we used 10 random partitions of held-in/held-out neurons, where 10% of neurons are held-out, and then reported the average across partitions. |
| Hardware Specification | No | The paper does not specify any particular hardware used for running the experiments, such as CPU, GPU models, or memory. |
| Software Dependencies | No | The paper does not provide specific version numbers for any software dependencies or libraries used in the implementation. |
| Experiment Setup | Yes | Concretely, the simulated circuit used three regions, each with a low-rank RNN architecture (Mastrogiuseppe & Ostojic, 2018; Beiran et al., 2023) so that low-dimensional dynamics could be easily visualized and compared. The activity in each region k evolves as y(k)t = (1 τk )y(k)t 1 + Wkϕ(y(k)t 1) + Xℓ =k Wk,ℓϕ(y(ℓ)t 1) + Gkc(k)t + ϵ(k)t where c(k)t is input to region k, read out linearly by Gk, Wk = Mk N k and Wk,ℓ= Mk,ℓN k,ℓ are low-rank within/between population weight matrices, of ranks 1 and 2, respectively, ϵt N(0, σ2I). Each RNN region had 128 neurons with tanh nonlinearities, and linear readouts for region-specific outputs. ... we then trained a three region, all-to-all connected MRDS-IR, with 2-dimensional local latent states, L1 = L2 = 2, order M = 1 filters for the channels, and a linear Gaussian observation model with diagonal noise. ... We then fit an MRDS-IR model using latent dimensions L1 = 3 and L2 = 2 and order M = 2 channels with a linear and Gaussian observation model with independent noise for each region s 128 neurons. |