Software and Code

Attention Regularization

Bootstrap-based regularization method to filter noisy attention scores and produce more interpretable explanations for vision transformers (ViT)

Class-Discriminative Attention Maps (CDAM)

Explainable AI method for vision transformer (ViT) to estimate the importance scores of input features with respect to the class or concept.

Concept Saliency Maps (CSM)

Evaluate and visualize latent representation of high-level concepts in generative models, such as variational autoencoders (VAEs).

Feature Perturbation Augmentation (FPA)

Data augmentation technique for deep learning training to reduce perturbation artifacts in downstream evaluation of explainable AI methods.

Jaccard - Similarity tests for binary data

Statistical tests of similarity between binary data using the Jaccard/Tanimoto coefficient – the ratio of intersection to union.

Jackstraw - Statistical Inference for Unsupervised Learning

Statistical methods to evaluate association between variables and their estimated latent variables. Latent variables may be estimated by principal component analysis (PCA), logistic factor analysis (LFA), clustering, and related techniques.

Tutorials

Association test with Principal Components
Statistical test of cluster memberships with the mtcars example
Unsupervised evaluation of cell identities in single cell genomics

Jaws - Jackstraw weighted shrinkage estimation

Jackstraw weighted shrinkage estimation for high-dimensional latent variable models. The jackstraw is used to estimate sparse loadings (i.e., coefficients) of Principal Component Analysis, Logistic Factor Analysis, and related techniques.

Obz AI

Obz AI is designed to bring explainability, continuous monitoring, and advanced outlier detection to AI-powered computer vision systems. With support for modern XAI methods, Obz AI enables ML engineers & scientists to ensure transparency, reliability, and trustworthiness in their vision models.