Leveraging unlabelled data for generalizable neural population decoding
Ximeng Mao, Nanda H. Krishna, Avery Hee-Woon Ryoo, Matthew G. Perich, Guillaume Lajoie
Read on arXiv →
Key claim
MOJO improves neural decoding with limited labeled data.
In plain English
Neurotechnologies like brain-computer interfaces rely on accurate neural decoders, but current models are limited by their dependence on labeled data. Existing spike-based models primarily use supervised learning, which restricts their training capabilities. The introduction of MOJO allows for a combination of self-supervised and supervised learning, improving performance, especially when labeled data is scarce. This advancement could enable builders to utilize unlabelled data more effectively, enhancing the scalability and flexibility of neural decoding applications.
Introduces a novel training framework combining self-supervised and supervised learning for spike-tokenizing models.
Evaluated on multiple datasets with clear performance improvements over existing methods.
Deep reliability assessment
The methodology supports improved performance in label-impoverished settings and generalization across tasks and species, but the claim of achieving performance comparable to neuro-foundation models for continuous signals may be overclaimed without direct comparison data.
Reproducibility
No open source code or dataset is mentioned in the paper.
Key figure
Figure 1 illustrates a POYO-style model augmented with MOJO, showing how latent representations from tokenized neural data are used for both supervised and self-supervised learning.