Tversky Neural Networks
8 days ago
- #Cognitive Psychology
- #Deep Learning
- #Neural Networks
- Introduction of Tversky Neural Networks, a new architecture changing the proximity function from scalar product to Tversky similarity.
- Tversky similarity is based on feature matching, accounting for common and distinctive features, aligning with human psychology.
- Differentiable parameterization of Tversky similarity allows integration into modern deep learning frameworks.
- New building blocks: Tversky Similarity Layer and Tversky Projection Layer, enhancing expressive capacity.
- Demonstrated improvements in image recognition and language modeling tasks.
- Tversky networks offer built-in interpretability, visualizing learned prototypes and features directly.
- Consistent learning of parameters confirming Tversky's salience hypothesis.