MOJO framework uses self-supervised learning to improve neural population decoding

Researchers introduce MOJO (Masked autOencoder-based JOint training), a framework for spike-tokenizing models that jointly leverages self-supervised learning via masked autoencoding and supervised learning objectives. This approach addresses the limitation of current models being restricted to supervised learning, which requires paired behavioral labels.

• Evaluated on three spiking datasets spanning monkey motor cortex and multi-regional mouse recordings, MOJO demonstrates superior performance over purely supervised-trained models.
• The improvement is particularly pronounced when training with limited labeled data, specifically in few-shot finetuning scenarios.
• Incorporating self-supervised learning yields more interpretable neuronal representations, improving performance on brain region classification and spike-statistics prediction without explicit optimization.
• MOJO generalizes to human electrocorticography during speech, achieving performance comparable to neuro-foundation models designed for continuous signals.

Augmenting spike-tokenizing models with self-supervised learning improves performance in label-poor settings and enables the use of unlabeled data across various tasks and species.