| name | mojo-ssl-neural-decoding |
| description | MOJO (Masked autOencoder-based JOint training) framework for decoding neural population activity using self-supervised learning with unlabelled data. Use when: working with limited labelled neural data, needing cross-session generalization, or wanting to leverage unlabelled spiking data for brain-computer interfaces. |
MOJO: Masked autOencoder-based JOint Training
Overview
MOJO is a training framework for spike-tokenizing neural models that jointly leverages self-supervised learning (SSL) via masked autoencoding and supervised learning (SL) objectives. This approach addresses the limitation of current spike-based models that are restricted to supervised learning, which limits training to datasets with paired behavioral labels.
Key Benefits
- Superior performance over purely SL-trained models, especially with limited labelled data
- Few-shot finetuning capability where only small amounts of labelled data from new sessions are available
- More interpretable neuronal representations improving brain region classification and spike-statistics prediction
- Cross-modality generalization beyond spiking data to human electrocorticography during speech
- Comparable performance to neuro-foundation models (NFMs) designed specifically for continuous signals
Implementation Steps
- Data Preparation: Tokenize neural data at the spike level to facilitate multi-session pretraining
- Model Architecture: Implement a transformer-based architecture capable of handling spike-tokenized sequences
- Joint Training Objective: Combine SSL via masked autoencoding with SL objectives:
- SSL: Randomly mask spike tokens and train model to reconstruct them
- SL: Train on available labelled behavioral data for decoding tasks
- Training Protocol:
- Pretrain on large unlabelled datasets using SSL objective
- Finetune on limited labelled data using combined SSL+SL objectives
- Evaluation: Test on multiple neural datasets (monkey motor cortex, mouse multi-regional recordings, human ECoG)
Use Cases
- Brain-computer interfaces with limited calibration data
- Cross-session neural decoding without extensive retraining
- Multi-species neural data analysis leveraging unlabelled recordings
- Neuro-foundation model pretraining for flexible downstream applications
Evaluation Datasets
- Monkey motor cortex during reaching tasks
- Multi-regional mouse recordings during vision and decision making tasks
- Human electrocorticography during speech production
Activation Keywords
neural decoding, spike-tokenizing, self-supervised learning, few-shot finetuning, brain-computer interface, neuro-foundation models, unlabelled neural data
References
- arXiv:2607.14086 - "Leveraging unlabelled data for generalizable neural population decoding"
- Subjects: Machine Learning (cs.LG), Neurons and Cognition (q-bio.NC)