Masked Autoencoders Decode Visual Perception from Spontaneous Brain Data at 84%

Clinical neuroprosthetics suffer from a data bottleneck: labeled perception trials are scarce while hours of spontaneous neural activity go unused. Researchers led by Aleksandr Kovalev tested whether self-supervised learning (SSL) could leverage this untapped resource. They pretrained a masked autoencoder on 14.6 hours of spontaneous multiunit activity recorded from an intracortical array implanted in a blind participant's primary visual cortex (V1). Remarkably, the model captured interpretable brain structure without any supervision: V1's spatial organization and perceptual state separation emerged purely from the latent representations, proving that resting-state activity is not random noise.

To evaluate the representations, the team used linear probing—training a logistic regression on the frozen latents—to decode perception from the same participant's stimulation trials. The model achieved 84.1% accuracy on a general psychometric task and 64.0% on a more difficult threshold-level detection task. This work demonstrates that unsupervised pretraining on spontaneous neural data is a promising strategy to overcome the labeling bottleneck in brain-computer interfaces. The paper was presented at the NeuroAI Workshop, AAAI 2026, and published in PMLR.