PNBA: Probabilistic Neural-Behavioral Representation Alignment [ICML2025]

Overview

PNBA addresses a fundamental question in neuroscience: how can neural circuits generate consistent function across individuals despite structural and physiological heterogeneity?

Our framework employs a probabilistic approach to align neural and behavioral representations across subjects through:

1. Probabilistic Representational Learning: Using distribution-level representation modeling to address neural variability across trials, sessions, and subjects.
2. Cross-modal Representational Alignment: Establishing reliable alignment between neural activity and behavior.
3. Generative Constraints: Preventing representation degeneration during the alignment process.

This approach reveals robust preserved neural representations across different brain regions and species, providing evidence for consistent neural coding principles despite anatomical differences.

Data

The data used in this project is available from the following repository: Sensorium 2023 Data

Usage

We here demonstrate our framework using an accessible example with mouse primary visual cortex (V1) data.

Training

We suggest the following hardware configurations:

• 8x NVIDIA A100 80GB GPUs
• 512GB RAM
• 2TB SSD storage for datasets and model checkpoints

To initiate training with the default configuration:

python train.py config/config.py

Inference and Evaluations

Before running inference:

1. Update the 'load_from' parameter in config/config.py with your trained model path. Here is the pretrained weights
2. Run the extraction script with appropriate GPU resources

# Extract neural representations (single GPU)
python extract_neural_rep.py config/config.py

# Verify the cross model alignment
python lv_statistics_neural_vs_stimulus.py

# Verify the zero-shot preserved neural presentation
python lv_statistics_activity.py

Acknowledgments

We would like to acknowledge:

• We acknowledge the neuralpredictors repository for providing valuable tools and infrastructure for neural data analysis that significantly contributed to this work.

• We thank the organizers and participants of the Sensorium2023 Competition for creating a robust benchmark and collaborative environment that helped advance research in neural representations.

• We acknowledge the foundational work by Safaie, Mostafa, et al. "Preserved neural dynamics across animals performing similar behaviour." Nature 623.7988 (2023): 765-771, which provided key insights into neural representational consistency on motor cortices.

Citation

If you use this code in your research, please cite our paper:

@article{zhu2025neural,
  title={Neural Representational Consistency Emerges from Probabilistic Neural-Behavioral Representation Alignment},
  author={Zhu, Yu, Song, Chunfeng, Ouyang, Wanli, Yu, Shan, Huang, Tiejun},
  journal={arXiv preprint arXiv:2505.04331},
  year={2025}
}