Drava

Drava aligns human cognitive concepts with semantic dimensions learned via disentangled representation learning (DRL) to improve interpretability of machine learning latent vectors used in data exploration such as t-SNE.

Key Features:

• Concept alignment and mismatch identification: Aligns user-defined human concepts with semantic dimensions derived from DRL and identifies mismatches between them.
• Interactive feedback mechanism: Incorporates iterative user feedback to reduce mismatches and refine the alignment between concepts and learned dimensions.
• Concept-driven exploration with visual piles: Supports concept-driven visual exploration using visual piles to refine concepts and inspect data guided by human-centric interpretations.
• Concept adaptor model: Employs a concept adaptor model to fine-tune DRL semantic dimensions based on user feedback.
• Addressing latent vector interpretability: Targets interpretability challenges of ML latent vectors commonly used in techniques such as t-SNE by aligning them with human concepts.

Scientific Applications:

• Interpretability of latent spaces: Enhances understanding of latent representations produced by DRL for downstream analysis and visualization.
• Concept-driven data exploration: Enables exploration of complex datasets guided by human concepts rather than solely by abstract latent dimensions.
• Visual analytics for researchers: Supports researchers and analysts in deriving insights from complex datasets through human-aligned semantic dimensions.

Methodology:

Derives semantic dimensions via disentangled representation learning (DRL), compares those dimensions to user-defined human concepts to identify mismatches, incorporates iterative user feedback, applies a concept adaptor model to fine-tune DRL semantic dimensions based on feedback, and uses visual piles for concept-driven visual exploration; the approach targets interpretability of latent vectors used in techniques such as t-SNE.