Self-Supervised Learning Decodes Perception from Spontaneous Neural Data
A study conducted by researchers at an unnamed institution explored the potential of self-supervised learning to utilize unlabeled spontaneous neural activity for enhancing perception decoding in clinical neuroprosthetics. They pretrained a masked autoencoder using 14.6 hours of spontaneous multiunit activity gathered from an intracortical array in the primary visual cortex (V1) of a blind participant. The model autonomously deciphered interpretable brain architecture, revealing V1's spatial arrangement and perceptual state distinctions from its latent representations. Performance was assessed via linear probing (logistic regression on frozen latents) with results showing an 84.1% accuracy on a general psychometric task and 64.0% on a challenging threshold-level task. This research indicates that spontaneous cortical activity holds structured information pertinent to perception, potentially solving the data scarcity issue in neuroprosthetics. The findings were shared on arXiv (ID: 2607.22615).
Key facts
- Masked autoencoder pretrained on 14.6 hours of spontaneous multiunit activity from V1 of a blind participant.
- Model captured V1's spatial organization and perceptual state separation without supervision.
- Linear probing achieved 84.1% accuracy on general psychometric task and 64.0% on threshold-level task.
- Study published on arXiv with ID 2607.22615.
- Spontaneous cortical activity shown to contain structured information relevant for perception.
- Self-supervised learning can leverage unlabeled neural data to improve perception decoding.
- Data bottleneck in clinical neuroprosthetics: labeled trials scarce, spontaneous activity underutilized.
- Intracortical array used for recording.
Entities
Institutions
- arXiv