Differentiable Tversky similarityThe key innovation is a differentiable parameterization of Tversky similarity. Next, Tversky neural networks are defined based on two new building blocks:Tversky Similarity Layer , analogous to metric similarity functions like dot product or cosine similarity. However, instead of simple scalar product, it uses a more complex and expressive nonlinear Tversky similarity function in the Tversky Similarity Layer. In total, the learnable parameters in Tversky neural networks include:prototype vectors, Πfeature vectors, Ωweights α, β, θΠ and Ω can be shared between different layers. And the learnable prototype and feature vectors in Tversky networks still need to be learned; not every initialization succeeded in the experiments.