SPHeRe

This is the official PyTorch implementation for "Rethinking Hebbian Principle: Low-Dimensional Structural Projection for Unsupervised Learning" (NeurIPS 2025)

Unsupervised learning

# Clone our repository
git clone https://github.com/brain-intelligence-lab/SPHeRe.git
cd SPHeRe

# Install and activate base anaconda environment
conda env create -n SPHeRe --file requirements.yml
conda activate SPHeRe

# Train a NN on CIFAR10 to reproduce the SOTA result printed on the paper
python main.py

# Or if you want to specify parameters:
python main.py [-h] [--cuda CUDA] [--seed SEED] [--dataset DATASET] [--is_bp]

For Tiny-ImageNet dataset, torchvision does not provide an api, you can use huggingface api, but because of internet instability, we choose to download manually and load the raw data from this repository.

Reconstruction experiment

To reimplement the AE for reconstruction, run:

python reconstruction/run.py --noise_mean 0.0 --noise_std 0.0

If you want to use your own pretrained models:

1. First run these to save the model files:

   python main.py
   python softhebb.py

2. The models will be saved in ./models/ folder

3. Then run the reconstruction code above

Citation

If you find this work useful, please cite our paper: