Research map / Label-efficient and robust learning / Self-supervised and contrastive

Leveraging unlabelled data for generalizable neural population decoding

NeurIPS 2026 · Label-efficient and robust learning › Self-supervised and contrastive · approach: Self-supervised neuroimaging representations · 0 citations (OpenAlex)

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 MOJO, a joint self-supervised (masked autoencoding of spike-rate latents) and supervised training framework for POYO-style spike-tokenizing neural decoders, improving decoding in label-scarce settings and transferring to ECoG.

Problem

Spike-tokenizing neural decoders (POYO family) are trained only with supervised learning, so they cannot use unlabelled neural recordings and require paired behavioural labels for pretraining.

Approach

Attach a masked-autoencoder SSL pathway on the input cross-attention latents of spike-tokenizing models: mask latents temporally and reconstruct unit spike counts with a Poisson NLL, jointly with the supervised behaviour loss, sharing encoder and backbone parameters.

Key results

• On monkey T-RT new-animal sessions, MOJO-POGRU (FT) reaches R2 0.7776 vs 0.7575 for POGRU (FT) and NDT-3 0.7066; MOJO-POYO (FT) gets 0.9131 on T-CO vs 0.8817 for POYO (FT).
• MOJO-POYO improves average balanced accuracy/R2 on IBL decision tasks to 0.756 vs 0.727 for POYO and 0.647 for NEDS.
• MOJO unit embeddings predict log mean firing rate with R2 0.88 vs 0.54 for pure-SL POYO.

Surprising findings

• Unit identification alone with MOJO-POSSM (<18K trainable parameters) matched fully finetuned supervised POSSM on unseen-animal T-RT sessions.
• Without enough data, a weakly learned SSL objective can be detrimental to SL performance.

Limitations

• Joint SSL-SL model is data-hungry; weak SSL can hurt SL performance with small data.
• Latent masking is temporal only; spatial masking of unseen neurons or regions is not addressed.
• Unit embeddings must be re-learned from scratch when finetuning to new sessions; ECoG needs modality-specific value-embedding layer.

Experimental setup

Datasets: Perich et al. monkey reaching (CO/RT) NLB Maze Allen Brain Observatory Neuropixels visual coding IBL Reproducible Electrophysiology human ECoG consonant-vowel syllable dataset

Benchmarks: monkey reaching decoding (R2) Allen mouse visual stimulus classification IBL mouse decision decoding human ECoG speech decoding brain region classification spike-statistic regression

Metrics: R2 accuracy balanced accuracy paired t-test p-values

Baselines: POYO POSSM (GRU) POSSM (Mamba) NDT-2 NDT-3 NEDS Du-IN EEGNet MLP GRU

Builds on: POYO POSSM masked autoencoder NDT-2 NEDS

Closest papers by meaning

• A Scalable, Causal, and Energy Efficient Framework for Neural Decoding with Spiking Neural Networks (NeurIPS 2025 · similarity 0.89)
• Towards a "Universal Translator" for Neural Dynamics at Single-Cell, Single-Spike Resolution (NeurIPS 2024 · similarity 0.88)
• Generalizable, real-time neural decoding with hybrid state-space models (NeurIPS 2025 · similarity 0.88)
• SPINT: Spatial Permutation-Invariant Neural Transformer for Consistent Intracortical Motor Decoding (NeurIPS 2025 · similarity 0.88)
• Neuroformer: Multimodal and Multitask Generative Pretraining for Brain Data (ICLR 2024 · similarity 0.88)
• The Brain's Bitter Lesson: Scaling Speech Decoding With Self-Supervised Learning (ICML 2025 · similarity 0.87)