Summary
A paper (arXiv 2607.14086) by Ximeng Mao, Nanda H. Krishna, Avery Hee-Woon Ryoo, Matthew G. Perich, and Guillaume Lajoie introduces MOJO (Masked autOencoder-based JOint training), a training framework for spike-tokenizing neural models that combines self-supervised learning (SSL) via masked autoencoding with supervised learning (SL) objectives. Current spike-based neural decoders, used in brain-computer interfaces and closed-loop experiments, are limited to supervised training on datasets with paired behavioral labels. MOJO removes this constraint by leveraging unlabelled neural data. Evaluated on three spiking datasets—monkey motor cortex recordings during reaching tasks and multi-regional mouse visual/decision-making recordings—MOJO outperforms purely supervised models, especially with limited labelled data and few-shot fine-tuning on new sessions. SSL also yields more interpretable neural representations, improving brain region classification and spike statistics prediction without explicit optimization. MOJO generalizes beyond spiking data to human speech cortex electrocorticography, matching neural foundation models designed for continuous signals. The results point toward more flexible, scalable data usage for training neural foundation models across tasks, species, and modalities.
Paper Overview
Field: Machine Learning
Authors: Ximeng Mao, Nanda H. Krishna, Avery Hee-Woon Ryoo, Matthew G. Perich, Guillaume Lajoie
Published: 2026-07-15
arXiv: 2607.14086
Abstract
Robust and accurate neural decoders are integral to neurotechnologies such as brain-computer interfaces and closed-loop experiments. Recent work has shown that tokenizing neural data at the spike level facilitates multi-session pretraining and delivers state-of-the-art decoding performance. However, current spike-based models are restricted to supervised learning (SL), limiting training to datasets with paired behavioural labels. To address this limitation, the authors introduce MOJO (Masked autOencoder-based JOint training), a training framework for spike-tokenizing models that jointly leverages self-supervised learning (SSL) via masked autoencoding and SL objectives.
Key contributions:
- Evaluated on three spiking datasets spanning monkey motor cortex during reaching tasks and multi-regional mouse recordings during visual/decision-making tasks.
- MOJO outperforms purely SL-trained models, with the largest gains when labelled data is limited — notably in few-shot fine-tuning where only small amounts of labelled data from new sessions are available.
- Incorporating SSL yields more interpretable neural representations, improving brain region classification and spike statistics prediction performance without explicit optimization for these tasks.
- MOJO generalizes beyond spiking data to human speech cortex electrocorticography (ECoG), continuing to outperform purely SL-trained models and achieving performance comparable to neural foundation models (NFMs) specifically designed for continuous signals.
Overall, augmenting spike-tokenizing models with SSL improves performance in label-scarce settings, enables the use of unlabelled data across tasks and species, and extends to other neural modalities — pointing toward more flexible and scalable data usage when training neural foundation models.
---
*Auto-collected on 2026-07-17*
This page is an English static mirror generated for search and AI citation.
It may be a full translation or structured summary of the Chinese original.
Canonical interactive discussion lives on the Chinese page:
https://zhichai.net/topic/178395196