tversky 0.2.0

Official Pytorch Library for Tversky Neural Networks

Official pytorch library for Tversky Neural Networks

Links

• pytorch library (this repository)
• ICLR 2026 Experiments
• ICLR 2026 Paper
• ICLR 2026 Page
• ICLR 2026 High Resolution Poster
• web site

Installation

Note: tversky requires PyTorch ≥ 2.0. Install it first following the instructions at https://pytorch.org/get-started. Then run:

pip install tversky

Notes

• The code used to reproduce the experiments presented in our ICLR 2026 paper is located in the tversky-networks-iclr2026 repository. This library was forked that repository. That repository does not use this library.

Component: Tversky Similarity Layer

from tversky import nn as tnn

sim_layer = tnn.TverskySimilarity(
    embedding_dim=64,
    fbank_size=128,
    similarity_model='contrast',
    normalize=False
)

Component: Tversky Projection Layer

from tversky import nn as tnn

proj_layer = tnn.TverskyProjection(
    embedding_dim=64,
    class_count=10,
    fbank_size=128,
    similarity_model='contrast',
    normalize=False
)

MNIST Example

make test-mnist

Results (plots, training curves, salience analysis): tests/outputs/mnist/report.md

License

LICENSE.txt

Citation

If you use this work, please cite the following paper:

@inproceedings{doumbouya2026tversky,
    title={Tversky Neural Networks: Psychologically Plausible Deep Learning with Differentiable Tversky Similarity},
    author={Moussa Koulako Bala Doumbouya and Dan Jurafsky and Christopher D Manning},
    booktitle={The Fourteenth International Conference on Learning Representations},
    year={2026},
    url={https://openreview.net/forum?id=koKWoKaMrE}
}