| name | masked-autoencoders-resting-state-neural-data |
| title | Masked Autoencoders Learn Perception-Relevant Representations from Resting State Neural Data |
| description | Self-supervised pretraining on spontaneous neural activity using masked autoencoders to improve perception decoding in clinical neuroprosthetics. Achieves 84.1% accuracy on psychometric tasks and 64.0% on threshold-level tasks. |
| arxiv_id | 2607.22615 |
| date | 2026-06-15T00:00:00.000Z |
| authors | ["Aleksandr Kovalev","Antonio Lozano","Fabrizio Grani","Cristina Soto Sanchez","Leili Soo","Rocío López-Peco","Adrian Villamarin-Ortiz","Roberto Morollón Ruiz","María del Mar Ayuso Arroyave","Alfonso Rodil","Eduardo Fernández"] |
| tags | ["neuroscience","self-supervised learning","masked autoencoders","neural decoding","clinical neuroprosthetics","resting state activity"] |
Masked Autoencoders for Resting State Neural Data
Overview
This methodology addresses the data bottleneck in clinical neuroprosthetics by leveraging spontaneous neural activity through self-supervised learning. The approach uses masked autoencoders pretrained on hours of unlabeled resting state neural data to improve perception decoding performance.
Key Contributions
1. Data Utilization Strategy
- Problem: Labeled perception trials are scarce while spontaneous neural activity is abundant but underutilized
- Solution: Self-supervised pretraining on 14.6 hours of spontaneous multiunit activity from intracortical V1 array
- Insight: Spontaneous cortical activity contains rich, task-relevant structure, not just noise
2. Emergent Brain Structure
- Spatial organization: V1's spatial organization emerges purely from latent representations without supervision
- Perceptual state separation: Clear separation of perceptual states in latent space
- Interpretability: Captures biologically meaningful brain structure through unsupervised learning
3. Performance Results
- General psychometric task: 84.1% perception decoding accuracy using linear probing
- Threshold-level task: 64.0% accuracy on more difficult discrimination task
- Method: Linear probing (logistic regression on frozen latents) with stimulation data
4. Clinical Impact
- Data efficiency: Reduces reliance on scarce labeled perception trials
- Transfer learning: Pretrained models can be fine-tuned for specific clinical applications
- Neuroprosthetic enhancement: Improves decoding accuracy for brain-computer interfaces
Implementation Guidelines
Model Architecture
class NeuralMaskedAutoencoder:
def __init__(self, input_dim, hidden_dim, mask_ratio=0.75):
self.encoder = TransformerEncoder(input_dim, hidden_dim)
self.decoder = TransformerDecoder(hidden_dim, input_dim)
self.mask_ratio = mask_ratio
def forward(self, neural_data):
masked_data, mask = self._apply_random_mask(neural_data, self.mask_ratio)
encoded = self.encoder(masked_data)
reconstructed = self.decoder(encoded)
return reconstructed, mask
def pretrain_on_resting_state(self, resting_state_data):
optimizer = AdamW(self.parameters())
for epoch in range(num_epochs):
loss = self._reconstruction_loss(resting_state_data)
loss.backward()
optimizer.step()
return self
def linear_probe_perception(self, stimulation_data, labels):
self.encoder.eval()
torch.no_grad():
features = .encoder(stimulation_data)
classifier = LogisticRegression()
classifier.fit(features, labels)
classifier
Training Protocol
-
Pretraining phase:
- Use 14.6+ hours of spontaneous multiunit activity
- Apply masked autoencoder objective (reconstruct masked neural activity)
- Train until convergence on reconstruction loss
-
Linear probing phase:
- Freeze pretrained encoder weights
- Extract features from stimulation data
- Train logistic regression classifier on extracted features
- Evaluate on perception decoding tasks
-
Evaluation metrics:
- General psychometric task accuracy
- Threshold-level task accuracy
- Comparison with supervised-only baselines
Applications
Clinical Neuroprosthetics
- Visual prosthetics: Improve decoding for blind participants with V1 implants
- Motor prosthetics: Extend to motor cortex for movement intention decoding
- Sensory restoration: Apply to other sensory modalities (auditory, somatosensory)
Neuroscience Research
- Brain structure discovery: Uncover latent organization in neural data
- State representation: Identify neural correlates of perceptual states
- Cross-species analysis: Apply to animal models for basic research
AI/ML Applications
- Self-supervised learning: General framework for neural data with limited labels
- Transfer learning: Pretrain on abundant unlabeled data, fine-tune on scarce labeled data
- Representation learning: Learn interpretable representations from neural time series
Limitations and Considerations
Data Requirements
- Requires substantial amounts of resting state neural data (>10 hours recommended)
- Quality of spontaneous activity affects pretraining effectiveness
- May need adaptation for different brain regions or recording modalities
Computational Resources
- Transformer-based architectures require significant memory
- Pretraining can be computationally expensive
- May need optimization for real-time applications
Clinical Translation
- Requires validation across multiple patients and conditions
- Regulatory approval needed for clinical deployment
- Integration with existing neuroprosthetic systems
Activation Keywords
Use this skill when working with:
- Neural decoding enhancement
- Self-supervised learning for neural data
- Resting state activity analysis
- Clinical neuroprosthetics
- Masked autoencoder pretraining
- Perception decoding accuracy
- Spontaneous neural activity utilization
References
- Primary: Kovalev, A., et al. (2026). Masked Autoencoders Learn Perception-Relevant Representations from Resting State Neural Data. arXiv:2607.22615 [q-bio.NC]. Proceedings of the First Workshop on NeuroAI Multimodal Intelligence @ AAAI 2026, PMLR 308:93-98.
- Related: Self-supervised learning in computer vision (MAE, BEiT)
- Applications: Clinical neuroprosthetics and brain-computer interfaces
Verification Steps
- Collect sufficient resting state neural data (>10 hours)
- Implement masked autoencoder architecture for neural time series
- Pretrain model on spontaneous activity with reconstruction objective
- Extract features from stimulation data using frozen encoder
- Train linear classifier and evaluate on perception decoding tasks
- Compare performance with supervised-only baseline approaches