Research map / Graph learning / Graph self-supervised learning
Self-Supervised Discovery of Neural Circuits in Spatially Patterned Neural Responses with Graph Neural Networks
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes a self-supervised graph neural network that jointly learns a latent connectivity matrix and predicts future spikes, evaluated on simulated ring attractor networks and real head direction cell recordings.
Problem
Inferring synaptic connectivity from partially observed neural population spike trains is hard, especially in strongly recurrent networks where spurious activity correlations among unconnected neurons mislead statistical inference methods.
Approach
A two-module architecture: a structure learning module (1D conv embeddings plus MLP) estimates pairwise weights, and a GNN spike prediction module with GRU message passing uses these weights to predict future Poisson spike rates; hidden neurons are added as interpolated auxiliary nodes so the model handles partial observation.
Key results
- On the fully observed threshold-crossing network, GNN achieves Delta 0.061 and L_bps 0.882 versus GLM 0.244 and 0.695.
- With localized excitation profile and external input, GNN reaches Delta 0.048 (threshold) versus GLM 0.125 and seqNMF 0.374.
- Shuffling the inferred interpolated weight rows raises inference error from 0.27 to 0.53, and real HD cell inference yields local excitation with surrounding inhibition consistent with ring attractor theory.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use GNN instead of GLM | Fully observed network, threshold-crossing · No external input | connectivity inference error Δ: 0.244 -> 0.061 | Helped |
| Use GNN instead of GLM | Fully observed network, LNP · No external input | connectivity inference error Δ: 0.238 -> 0.049 | Helped |
| Use GNN instead of GLM | Fully observed network, threshold-crossing · External input | connectivity inference error Δ: 0.259 -> 0.073 | Helped |
| Use GNN instead of GLM | Fully observed network, LNP · External input | connectivity inference error Δ: 0.245 -> 0.058 | Helped |
| Use GNN instead of GLM | Fully observed network, threshold-crossing · Weakly correlated, with external input | connectivity inference error Δ: 0.125 -> 0.048 | Helped |
| Use GNN instead of GLM | Fully observed network, LNP · Weakly correlated, with external input | connectivity inference error Δ: 0.117 -> 0.043 | Helped |
| Add structured external input | Fully observed network · Synthetic ring network | spike prediction L_bps | Helped |
| Add structured external input | Fully observed network · Synthetic ring network | connectivity inference error Δ | Mixed |
| Change Mexican-hat connectivity to localized excitation with broadly tuned inhibition | Fully observed network with external input · Weakly correlated activity | connectivity inference error Δ | Helped |
| Increase the number of observed neurons | Partially observed GNN network · Simulated HD network | connectivity inference error Δ | Helped |
| Add hidden neurons | Partially observed GNN network · Simulated HD network | spike prediction accuracy | Helped |
| Increase the hidden population from 5 to 20 at a fixed observed-neuron count | Partially observed GNN network · Simulated HD network | connectivity inference error Δ | Mixed |
Surprising findings
- Expanding hidden neurons yields diminishing returns in spike prediction and can plateau or raise inference error, indicating a trade-off between prediction and structural inference.
- External input improved spike predictability but did not reduce connectivity inference error, which stayed similar or slightly elevated.
Limitations
- Assumes underlying connectivity remains fixed over the observation period; time-varying plasticity not modeled.
- Evaluated only on 1D ring attractor geometry and a single 19-cell HD session; inferred connectivity is not unique.
- Real-data inferred weights show variability attributed to limited observed fraction.
Experimental setup
Datasets: synthetic ring attractor spike data (100 neurons) CRCNS head direction cell benchmark (19 HD cells, ADn, freely moving mice)
Benchmarks: synthetic ring attractor networks (threshold-crossing and LNP) CRCNS HD cell dataset
Metrics: normalized inference error Delta bits per spike L_bps
Baselines: GLM seqNMF TCA
Builds on: generalized linear model tensor component analysis seqNMF NetFormer AMAG
Closest papers by meaning
- NetFormer: An interpretable model for recovering dynamical connectivity in neuronal population dynamics (ICLR 2025 · similarity 0.89)
- Partial observation can induce mechanistic mismatches in data-constrained models of neural dynamics (NeurIPS 2024 · similarity 0.87)
- SGNNBench: A Holistic Evaluation of Spiking Graph Neural Networks on Large-scale Graphs (NeurIPS 2026 · similarity 0.87)
- Dynamic Reactive Spiking Graph Neural Network (AAAI 2024 · similarity 0.87)
- Autaptic Synaptic Circuit Enhances Spatio-temporal Predictive Learning of Spiking Neural Networks (ICML 2024 · similarity 0.86)
- Temporal Spiking Neural Networks with Synaptic Delay for Graph Reasoning (ICML 2024 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.