Research map / Label-efficient and robust learning / Zero-/few-shot and meta-learning

SpikeSSL: A Universal Spike Inference Framework with Dynamics-Informed State-Space Layers

NeurIPS 2026 · Label-efficient and robust learning › Zero-/few-shot and meta-learning · approach: Zero-shot robustness and transfer

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 SpikeSSL, a bidirectional IIR state-space network with multi-modal AdaLN conditioning and heteroscedastic uncertainty for spike inference from calcium imaging, augmented with a biophysical simulator, reporting SOTA in-domain and zero-shot leave-one-indicator-out results.

Problem

Spike inference from two-photon calcium fluorescence generalizes poorly across calcium indicators with different kinetics and signal statistics, so supervised models degrade on unseen indicators.

Approach

A temporal backbone of bidirectional first-order IIR scans with a multi-scale decay basis, conditioned on indicator identity, sampling rate, and trace statistics via AdaLN and decay-logit shifts, plus a heteroscedastic variance head and a biophysically varied calcium simulator for augmentation.

Key results

• In-domain mean correlation 0.741 for SpikeSSL vs 0.545 for ENS2 and 0.524 for CASCADE.
• Zero-shot LOIO mean correlation 0.654 for SpikeSSL vs 0.483 for ENS2 and 0.460 for CASCADE.
• With simulation augmentation, LOIO mean correlation rises to 0.704 for SpikeSSL + Sim.

Surprising findings

• Scaling SpikeSSL to a larger variant did not yield additional gains, while scaling baselines improved them but not to SpikeSSL's level.
• SpikeSSL's zero-shot error on the OGB-1 split (1.979) is worse than CASCADE (1.208) and ENS2 (1.005), despite higher correlation.

Limitations

• Calibration analysis limited to the five-split LOIO benchmark; behavior under more extreme acquisition shifts is untested.
• Zero-shot error and bias are worse than some baselines on certain splits (e.g., OGB-1 error, SC-GCaMP6s bias).

Experimental setup

Datasets: ground-truth calcium-electrophysiology database (33 datasets, 1,960 neurons) simulated calcium traces (~11,000)

Benchmarks: five fixed in-domain and LOIO splits from the 33-dataset ground-truth database

Metrics: Pearson correlation normalized error mean bias

Baselines: OASIS MLSpike CASCADE ENS2

Builds on: OASIS MLSpike CASCADE ENS2 beta-NLL

Closest papers by meaning

• Active learning of neural population dynamics using two-photon holographic optogenetics (NeurIPS 2024 · similarity 0.85)
• P-SPIKESSM: HARNESSING PROBABILISTIC SPIKING STATE SPACE MODELS FOR LONG-RANGE DEPENDENCY TASKS (ICLR 2025 · similarity 0.85)
• Towards a "Universal Translator" for Neural Dynamics at Single-Cell, Single-Spike Resolution (NeurIPS 2024 · similarity 0.85)
• Leveraging unlabelled data for generalizable neural population decoding (NeurIPS 2026 · similarity 0.85)
• SimSort: A Data-Driven Framework for Spike Sorting by Large-Scale Electrophysiology Simulation (NeurIPS 2025 · similarity 0.85)
• Adaptive Calibration: A Unified Conversion Framework of Spiking Neural Networks (AAAI 2025 · similarity 0.84)