Browse Teach Curate

Categorical Graphical Language for Compositional Deep-Learning Training Regimes

The learning diagram: a graphical language that represents a deep-learning training setup - its datasets, component models, and their interactions - as structured data that compiles to a single loss function whose minimization makes the components' predictions agree. Given a rigorous category-theoretic semantics, it unifies setups such as knowledge distillation, multi-task and multi-modal learning and supports composing, relating and editing models on a firm mathematical footing.

Demerzel Text

INTRODUCTION DIAGRAMMATICLEARNING: A GRAPHICAL LAN- GUAGE FOR COMPOSITIONAL TRAINING REGIMES Mason

This paper introduces the learning diagram, a graphical language that represents a deep-learning training setup - its datasets, component models, and the ways they interact - as structured data rathe…