leveraging-unlabelled-neural-decoding

name	leveraging-unlabelled-neural-decoding
description	Skill for understanding and applying the MOJO framework for leveraging unlabeled data in neural population decoding via masked autoencoding and joint supervised learning.
activation	leveraging unlabelled neural decoding, MOJO framework, SSL for neural decoding

Leveraging Unlabelled Data for Generalizable Neural Population Decoding

Overview

This skill summarizes the MOJO (Masked autoencoder-based Joint training) framework introduced in arXiv:2607.14086v1. MOJO combines self-supervised learning (SSL) via masked autoencoding with supervised learning (SL) to train spike-tokenizing models for neural population decoding. It enables the use of unlabeled neural data across tasks and species, improving performance in label-scarce scenarios and yielding more interpretable neuronal representations.

Core Concepts

• Spike-tokenizing models: Convert raw spike trains into discrete token sequences, enabling sequence modeling techniques.
• Masked Autoencoding (MAE): A self-supervised objective where random tokens are masked and the model learns to reconstruct them, capturing rich spatiotemporal dependencies.
• Joint Training: MOJO optimizes a combined loss: L = λ * L_SSL + (1-λ) * L_SL, balancing representation learning and task-specific prediction.
• Generalization: By leveraging unlabeled data, MOJO learns robust features that transfer across sessions, subjects, and even modalities (e.g., to human ECoG).

Methodology

1. Tokenization: Convert spike trains into discrete tokens using a learned vector-quantized variational autoencoder (VQ-VAE) or similar discretization.
2. Masking: Randomly mask a proportion (e.g., 15-40%) of tokens in the input sequence.
3. SSL Objective: Predict the masked tokens using an autoregressive or bidirectional transformer encoder-decoder.
4. SL Objective: Predict behavioral labels (e.g., movement direction, stimulus identity) from the latent representations.
5. Optimization: Jointly minimize the weighted sum of SSL and SL losses.
6. Evaluation: Assess decoding performance, robustness to limited labels, and interpretability via auxiliary tasks (brain region classification, spike-statistics prediction).

Steps to Apply

1. Prepare Spike Data: Sort and bin spike trains into time bins (e.g., 10-50 ms) to create binned spike count matrices.
2. Tokenize: Train a tokenizer (e.g., VQ-VAE) on binned spike data to map sequences to discrete token IDs.
3. Mask Tokens: Randomly mask tokens in the tokenized sequences during training.
4. Build Model: Use a transformer-based architecture (encoder for context, decoder for reconstruction; optionally a prediction head for SL).
5. Define Losses:
   • SSL: Cross-entropy between predicted and actual masked tokens.
   • SL: Task-appropriate loss (e.g., cross-entropy for classification, MSE for regression).
   • Combined: L = α * L_SSL + (1-α) * L_SL (α typically 0.5).
6. Train: Optimize end-to-end on labeled and unlabeled data batches.
7. Evaluate: Test decoding accuracy on held-out labeled sessions; probe representational quality via downstream tasks.

Pitfalls

• Tokenization Quality: Poor tokenization limits downstream performance; ensure the tokenizer captures sufficient spike-train variability.
• Masking Ratio: Too high masking reduces SSL signal; too low reduces regularization. Tune via validation.
• Balancing α: Imbalanced weighting can cause one objective to dominate; monitor both losses.
• Generalization Gap: Validate that SSL pretraining indeed improves few-shot adaptation, not just overall accuracy.
• Modality Shift: When applying to new modalities (e.g., ECoG), re-tokenize or fine-tune the tokenizer.

References

• Primary paper: X. Mao et al., "Leveraging unlabelled data for generalizable neural population decoding", arXiv:2607.14086v1, 2026.
• Related works: Masked Autoencoders for vision (He et al., 2021), SpikeGPT (unknown), Neural Decoding literature.

Activation Examples

• leveraging unlabelled neural decoding
• MOJO framework neural decoding
• SSL for spike-tokenizing models