Get the latest tech news
Tversky Neural Networks
Psychologically Plausible Deep Learning with Differentiable Tversky Similarity
Despite its psychological plausibility, this model relied on discrete set operations, making it incompatible with the differentiable, gradient-based optimization that underlies modern deep learning. Image Recognition: When adapting frozen ResNet-50 for classification tasks, using a Tversky projection layer instead of linear at the output (=Tversky-ResNet-50) led to accuracy improvements from 36.0% to 44.9% (NABirds) and from 57.4% to 62.3% (MNIST). Its fundamental operations are based on psychologically intuitive concepts of common and distinctive features, providing a built-in, transparent language for explaining model reasoning.
Or read this on Hacker News