Research map / Probabilistic and scientific modeling / Dynamical systems
convSeq: Fast and Scalable Method for Detecting Patterns in Spike Data
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.
Introduces convSeq, a backpropagation-optimized constrained 2D convolutional filter method for unsupervised detection of spike sequences in neural recordings, reported to be over 100x faster than seqNMF and PP-Seq.
Problem
Detecting repeating spatiotemporal spike sequences in large neural recordings without behavioral references is computationally hard; existing methods (convNMF, seqNMF, PP-Seq) do not scale well to large datasets.
Approach
Optimize a set of constrained 2D filters with backpropagation (softmax-normalized filters with variance-maximization, total variation, and cross-correlation penalties), optionally parameterized as truncated Gaussians with learnable means, and assess peak significance via a null distribution of random filters.
Key results
- On the largest dataset (500000 timesteps), convSeq is about 32 times faster than PP-Seq on CPU, and GPU further reduces run time by about six times.
- convSeq outperforms seqNMF and PP-Seq in detection performance across all 81 synthetic grid conditions tested.
- convSeq correctly disentangles forward and reverse bidirectional sequences and two partially overlapping sequences on synthetic data, and extracts place cell sequences from mouse CA1 data.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use convSeq instead of PP-Seq or seqNMF | Synthetic sequence-detection datasets · 81 synthetic datasets | detection performance | Helped |
| use convSeq instead of PP-Seq | Sequence detection · Largest speed-benchmark dataset (500000 timesteps) | speed: about 32 times faster | Helped |
| enable GPU (ablation) | convSeq · Speed benchmark | runtime: a factor of six | Helped |
| use parameterized truncated Gaussian filters instead of direct filter optimization (ablation) | convSeq · Synthetic and real-data experiments | detection performance | No clear effect |
| reduce the total-variation weight to 1.5 (ablation) | convSeq · Synthetic sequence detection | false positives | Hurt |
| increase the total-variation weight to 100 (ablation) | convSeq · Synthetic sequence detection | false positives | Helped |
Surprising findings
- Two partially overlapping sequences produce undesirable sub-threshold peaks in the other pattern's filter response, showing unidirectional patterns with shared neurons are hard to disentangle cleanly.
Limitations
- Overlapping unidirectional sequences produce sub-threshold spurious peaks; detailed handling left to future work.
- Significant time warping (durations varying by more than 2x) requires running with multiple filter widths M.
- Choice of K and M is empirical and requires search; the method does not infer them automatically.
Experimental setup
Datasets: synthetic spike sequence datasets built from mouse CA1 recording background mouse CA1 recording (island dataset, zivlab) songbird HVC recording simulated 2D place cell dataset
Benchmarks: synthetic spike sequence grid (81 datasets) run time scaling on synthetic datasets mouse CA1 place cell sequences songbird HVC sequences
Metrics: true positive rate false positive rate false negative rate ROC curves run time
Baselines: seqNMF PP-Seq
Builds on: convNMF seqNMF
Closest papers by meaning
- Enabling hyperparameter optimization in sequential autoencoders for spiking neural data (NeurIPS 2019 · similarity 0.86)
- Spike-based Digital Brain: a novel fundamental model for brain activity analysis (ICLR 2026 · similarity 0.85)
- I2E: Real-Time Image-to-Event Conversion for High-Performance Spiking Neural Networks (AAAI 2026 · similarity 0.85)
- SimSort: A Data-Driven Framework for Spike Sorting by Large-Scale Electrophysiology Simulation (NeurIPS 2025 · similarity 0.84)
- CREIMBO: Cross-Regional Ensemble Interactions in Multi-view Brain Observations (ICLR 2025 · similarity 0.83)
- A Scalable, Causal, and Energy Efficient Framework for Neural Decoding with Spiking Neural Networks (NeurIPS 2025 · similarity 0.83)
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.