| name | leveraging-unlabelled-data-for-generalizable-neural-population-decoding |
| description | Skill for understanding and applying the research from arXiv:2607.14086 "Leveraging unlabelled data for generalizable neural population decoding" |
| category | ai_collection |
leveraging-unlabelled-data-for-generalizable-neural-population-decoding
Paper Information
- Title: Leveraging unlabelled data for generalizable neural population decoding
- arXiv ID: 2607.14086
- Authors: Ximeng Mao, Nanda H. Krishna, Avery Hee-Woon Ryoo, Matthew G. Perich, Guillaume Lajoie
- Published: 2026-07-16 (based on arXiv listing)
- Subjects: Machine Learning (cs.LG); Neurons and Cognition (q-bio.NC)
- Comments: Not specified in the arXiv listing
Core Concepts
MOJO Framework
The paper introduces MOJO (Masked autoEncoder-based Joint training), a training framework for spike-tokenizing models that jointly leverages self-supervised learning (SSL) via masked autoencoding and supervised learning (SL) objectives.
Key Findings
- Current spike-based models are restricted to supervised learning, limiting training to datasets with paired behavioral labels
- MOJO combines SSL (via masked autoencoding) with SL objectives to improve decoding performance
- MOJO demonstrates superior performance over purely SL-trained models, especially with limited labeled data
- SSL yields more interpretable neuronal representations, improving performance on brain region classification and spike-statistics prediction
- MOJO generalizes beyond spiking data to human electrocorticography during speech, achieving performance comparable to neuro-foundation models
- Augmenting spike-tokenizing models with SSL improves performance in label-impoverished settings and enables use of unlabelled data across tasks and species
Technical Approach
- Spike Tokenization: Converting neural spike trains into discrete tokens for modeling
- Masked Autoencoding: Self-supervised objective where portions of the spike token sequence are masked and the model predicts them
- Joint Training: Combining the SSL reconstruction loss with supervised decoding loss
- Cross-modal Generalization: Demonstrating that the learned representations transfer to other neural modalities (e.g., ECoG)
Applications in Agent Design
- Robust Neural Decoding: Build more resilient brain-computer interface decoders that work with limited labeled data
- Self-supervised Pretraining: Use large amounts of unlabeled neural data to pretrain decoding models
- Cross-task Generalization: Develop neural representations that transfer across different behavioral tasks and species
- Interpretable Representations: Learn neural representations that are meaningful for understanding brain function
- Multi-modal Integration: Combine spike-based models with other neural signal types (EEG, ECoG, fMRI) through shared latent spaces
Implementation Guidelines
For implementing MOJO-like approaches in neural decoding systems:
- Tokenization Pipeline: Convert neural spike trains into discrete token sequences using methods like spike sorting or feature extraction
- Masking Strategy: Implement random masking of tokens during training (typically 15-50% masking rate)
- Autoencoder Architecture: Use transformer-based or recurrent architectures capable of sequence modeling
- Joint Loss Function: Combine reconstruction loss (SSL) with prediction loss (SL) using appropriate weighting
- Pre-training Strategy: Pre-train on large unlabeled datasets, then fine-tune on limited labeled data for specific tasks
- Evaluation Metrics: Measure decoding accuracy, representation interpretability, and cross-task generalization
Activation Keywords
unlabelled data, neural population decoding, MOJO, masked autoencoding, self-supervised learning, spike-tokenizing models, brain-computer interface, neural representation learning, computational neuroscience
References
- arXiv:2607.14086 - Leveraging unlabelled data for generalizable neural population decoding
- Related work on self-supervised learning for neural data, masked autoencoding, and brain-computer interfaces